Code & Community with Copilot CLI - CourtneyR.dev's 2nd Brain

Code & Community with Copilot CLI 🔗 **Source:** [Code & Community with Copilot CLI](https://www.youtube.com/watch?v=bRtK041JFFE) > [!note]+ Document Notes > tags: #github-copilot #cli #agents #mcp #developers > author: GitHub > source: YouTube > reference: https://www.youtube.com/watch?v=bRtK041JFFE > date created: 2025-01-13 08:00 > > # 🏷️ Title: Code & Community with Copilot CLI >  ## 🧰 Action Inbox - [H1] Try fleet mode with autopilot for complex tasks — demonstrated live #github-copilot - [H2] Set up WorkIQ plugin for Microsoft 365 integration — improves productivity #workflow-automation - [H3] Use chronicle tips to analyze CLI usage patterns — self-improvement tool #productivity ## 🧭 Table of Contents - [[#📝 Summary]] - [[#📖 Main Content]] - [[#🗒️ Notes]] - [[#📚 References]] - [[#🗣️ Paraphrase]] - [[#📰 Newsletter]] - [[#✅ Action Items]] - [[#🧪 Hypotheses]] - [[#📊 Methodology]] - [[#❓ Questions]] - [[#🔮 Future Research]] - [[#💡 Implications]] - [[#🧠 Learning Styles Assessment]] - [[#📊 Accessibility Assessment]] - [[#🌍 DEIB Assessment]] - [[#✅ Logical Validity]] - [[#🛡️ DARVO Analysis]] - [[#🎯 Conclusion]] ### 🔗 References - [GitHub Copilot CLI](https://cli.github.com) - [Awesome Copilot Plugins](https://github.com/github/awesome-copilot) ## 📝 Summary - GitHub's Product Manager Ryan Hedge demonstrates the dramatic evolution of GitHub Copilot CLI since its public preview - CLI has grown from basic commands to a comprehensive agentic harness with 167 releases in ~5 months - New features include fleet mode (parallel agents), WorkIQ integration, chronicle analysis, and extensive plugin ecosystem - Tool supports headless operation, multiple model choices, and deep customization through skills, agents, and MCP servers - Bottom line: Copilot CLI has become a complete terminal-based AI assistant for coding, productivity, and knowledge work ## 📖 Main Content The GitHub Copilot CLI has undergone explosive development since its September 2024 public preview, with the team shipping new features daily. Ryan Hedge, Product Manager at GitHub, showcased how the tool has evolved from a simple command interface into a sophisticated agentic platform. The CLI now supports multiple AI models (Claude, GPT, Gemini) through a single subscription, allowing developers to dispatch different models for the same task and compare results. This model flexibility eliminates the need for multiple AI subscriptions. Key architectural advances include fleet mode, which automatically breaks complex tasks into parallel subtasks executed by multiple agents simultaneously. Combined with autopilot mode, this creates fully autonomous workflows where agents work collaboratively to complete objectives. The platform's extensibility comes through three main mechanisms: skills (simple text instructions), MCP servers (Model Context Protocol integrations), and custom agents. The community-driven awesome-copilot repository provides hundreds of pre-built extensions. Integration capabilities extend beyond coding to knowledge work through plugins like WorkIQ, which connects Microsoft 365 services, allowing natural language queries across email, calendar, and documents directly from the terminal. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/bRtK041JFFE?t=0) — [H1] Introduction and evolution overview - Key idea: Copilot CLI has become host's primary development tool, replacing traditional editors - Evidence: "I don't think I've opened insiders like I think the last time I opened it was a couple of weeks ago" #### ⏱️ [02:30](https://youtu.be/bRtK041JFFE?t=150) — [H2] Explosive development pace shown through releases - Key idea: 167 releases in ~5 months with daily feature additions - Evidence: "17 pages of releases... some days we ship more than once" #### ⏱️ [05:45](https://youtu.be/bRtK041JFFE?t=345) — [H2] Terminal AI reintroduces power of command line - Key idea: Terminal-based AI assistance shows developers full computer capabilities - Evidence: "The rise of terminal based AI assistance has reintroduced people to just how powerful the terminal is" #### ⏱️ [08:20](https://youtu.be/bRtK041JFFE?t=500) — [H1] Multi-model support eliminates subscription sprawl - Key idea: Single subscription provides access to multiple leading AI models - Evidence: "Being able to when a new model ships know that you have one subscription that's going to get you access to all of them, that's huge" #### ⏱️ [11:15](https://youtu.be/bRtK041JFFE?t=675) — [H1] Parallel model dispatch for better results - Key idea: Can run same task across multiple models simultaneously for comparison - Evidence: "Spin up parallel agents using Opus, using codeex, and using Gemini and have each of them review the plan" #### ⏱️ [14:30](https://youtu.be/bRtK041JFFE?t=870) — [H2] Headless operation enables automation integration - Key idea: CLI works in scripts and GitHub Actions for automated workflows - Evidence: "We will utilize this to help triage issues that come in... in a GitHub action" #### ⏱️ [18:45](https://youtu.be/bRtK041JFFE?t=1125) — [H2] Knowledge work automation beyond coding - Key idea: CLI automates repetitive business tasks across multiple systems - Evidence: Demonstrated Azure data extraction to Google Sheets workflow #### ⏱️ [22:10](https://youtu.be/bRtK041JFFE?t=1330) — [H1] WorkIQ plugin integrates Microsoft 365 ecosystem - Key idea: Natural language queries across email, calendar, and documents - Evidence: "What meetings do I have going on next week? What documents have I been pinged in?" #### ⏱️ [26:30](https://youtu.be/bRtK041JFFE?t=1590) — [H2] Community plugin ecosystem through awesome-copilot - Key idea: Hundreds of community-contributed plugins available for installation - Evidence: Extensive skills folder with curated, tested community contributions #### ⏱️ [29:45](https://youtu.be/bRtK041JFFE?t=1785) — [H2] Skills are simple text instructions for agent behavior - Key idea: Creating custom skills requires only name, description, and plain English instructions - Evidence: "All you have to do to write a skill is set a name for it, give a description for it... you just literally describe what you want it to do" #### ⏱️ [34:20](https://youtu.be/bRtK041JFFE?t=2060) — [H1] Chronicle feature provides usage analytics and coaching - Key idea: AI analyzes your CLI usage patterns to suggest improvements - Evidence: "/chronicle tips" analyzes conversation history to optimize workflows #### ⏱️ [38:50](https://youtu.be/bRtK041JFFE?t=2330) — [H3] Control+G opens prompts in default editor - Key idea: Long prompts can be edited in familiar editor environment - Evidence: Demonstrated Vim integration for complex prompt editing #### ⏱️ [42:15](https://youtu.be/bRtK041JFFE?t=2535) — [H2] VS Code integration provides contextual awareness - Key idea: CLI running in VS Code terminal accesses open files and project context - Evidence: "Visual Studio Code Insiders connected" message and context integration #### ⏱️ [46:30](https://youtu.be/bRtK041JFFE?t=2790) — [H1] Fleet mode enables parallel agent execution - Key idea: Complex tasks automatically split into parallel subtasks with dependency management - Evidence: "Create a to-do list... figure out which ones can be done independently... dispatch several parallel agents" #### ⏱️ [50:45](https://youtu.be/bRtK041JFFE?t=3045) — [H1] Autopilot + Fleet creates fully autonomous workflows - Key idea: Combined modes create self-directing agent fleets that complete entire features - Evidence: "Make a plan, get the plan... slash fleet implement the plan. And then it just goes off and does that" ## 📚 References - [GitHub Copilot CLI Documentation](https://docs.github.com/en/copilot/copilot-cli) - [Awesome Copilot Repository](https://github.com/github/awesome-copilot) - [Agent Skills Specification](https://agentskills.io) - [WorkIQ Plugin Documentation](https://github.com/microsoft/copilot-cli-plugin-workiq) ## 🗣️ Paraphrase - GitHub Copilot CLI has transformed from a simple command interface into a comprehensive AI development platform with rapid daily feature additions and community contributions - The tool eliminates the need for multiple AI subscriptions by providing access to all major models through one interface, enabling parallel model comparison and task execution - Beyond coding, the CLI serves as a universal automation platform for knowledge work, integrating with business systems through plugins and enabling fully autonomous agent workflows ## 📰 Newsletter **Copilot CLI: The Terminal Revolution for Developers** GitHub's Copilot CLI has quietly become one of the most powerful AI tools for developers, evolving far beyond its original scope. With 167 releases in just five months, this terminal-based assistant now offers features that rival dedicated AI applications. **Key Benefits:** - **One subscription, all models**: Access Claude, GPT, and Gemini through a single interface - **Parallel processing**: Fleet mode automatically breaks tasks into concurrent subtasks - **Beyond coding**: Automate knowledge work with Microsoft 365 integration - **Infinite customization**: Hundreds of community plugins and simple skill creation The CLI's strength lies in combining the terminal's power with modern AI capabilities, creating workflows impossible in traditional editors. Try `/chronicle tips` to analyze your usage patterns—you might be surprised what you learn about your own development habits. **Getting Started:** - Install with your existing GitHub Copilot subscription - Explore the [awesome-copilot](https://github.com/github/awesome-copilot) repository for plugins - Learn skill creation at [agentskills.io](https://agentskills.io) Resources: [GitHub CLI](https://cli.github.com), [Copilot Documentation](https://docs.github.com/en/copilot), [Agent Skills](https://agentskills.io) ## ✅ Action Items #### DevRel - Developer advocacy: Demo fleet mode capabilities at conferences, create comparison content showing multi-model workflows - Education: Develop workshops on CLI automation, create skill-writing tutorials for different experience levels - Developer experience: Document MCP vs skills decision matrix, improve onboarding for non-coding use cases - Developer marketing: Highlight productivity gains from terminal-first development, showcase community plugin ecosystem - Community management: Curate awesome-copilot repository, facilitate skill sharing between developers #### OSPO - Policy & Governance: Review plugin security model, establish community contribution guidelines for awesome-copilot - License Compliance: Audit community plugin licenses, ensure skill-sharing compliance - Security Management: Assess headless operation security, review MCP server trust boundaries - Software Inventory: Track CLI plugin dependencies, maintain SBOM for integrated tools - Developer Support: Create internal CLI adoption guides, establish best practices for enterprise deployment - Community Engagement: Support plugin developer onboarding, fund high-impact community tools - Strategic Alignment: Measure CLI adoption impact on development velocity, track multi-model usage patterns #### Community - Maintainers: Improve plugin discovery mechanisms, streamline skill contribution process - Contributors: Create first-plugin templates, establish skill quality guidelines - Sponsors: Fund plugin development sprints, support community skill competitions - Ecosystem: Coordinate with editor extension teams, align on cross-platform plugin standards #### WordPress - Core: Explore CLI integration for WordPress development workflows, create WordPress-specific skills - Core Performance: Use CLI for automated performance regression detection, create profiling skills - Core AI: Develop WordPress AI guidelines skills, create evaluation automation - Community: Create WordCamp presentation templates about CLI usage, develop local development skills - Training: Build CLI-based WordPress learning paths, create interactive tutorial skills - Documentation: Generate DevHub content using CLI agents, create documentation review workflows - CLI: Develop WP-CLI + Copilot CLI integration patterns, create WordPress deployment skills #### The WPCC - Programs: Use CLI for fellowship project automation, create grant application review workflows - Projects: Develop governance documentation generation, automate project health checks - Funding: Create transparent funding decision workflows, automate impact reporting #### FAIR - Technical independence: Verify CLI multi-model access remains vendor-neutral, audit plugin marketplace governance - Policies: Ensure skill-sharing promotes ecosystem diversity, prevent vendor lock-in through plugin standards - Governance artifacts: Document CLI governance model, establish community plugin policies ## 🧪 Hypotheses - Developers using fleet mode with multiple models produce higher quality code than single-model workflows due to built-in review processes - Terminal-first AI development increases productivity more than GUI-based alternatives because it reduces context switching - Community-contributed skills create better domain-specific solutions than vendor-provided features due to specialized knowledge - Chronicle-driven usage analysis leads to measurable improvements in developer AI interaction patterns ### 📊 Methodology - Study design to test hypothesis 1: Compare code quality metrics across fleet vs single-agent implementations over 30-day periods - Data sources and analysis for hypothesis 2: Track context switch frequency and task completion times between CLI and GUI AI tools - Experiments to validate hypothesis 3: Measure effectiveness of community vs vendor skills across standardized development tasks ### ❓ Questions - How does fleet mode performance scale with task complexity and what are the optimal parallelization strategies? - What governance models work best for community plugin ecosystems while maintaining security and quality? - How can chronicle analytics be used to automatically optimize individual developer AI interaction patterns? ### 🔮 Future Research - Multi-agent coordination patterns for complex software development projects requiring specialized domain knowledge - Integration strategies between terminal AI and traditional development environments that preserve both workflows - Community-driven AI skill development models that could apply beyond development tools to other professional domains ### 💡 Implications - Practice: Terminal-first AI development may fundamentally change how developers interact with their tools and organize their workflows - Policy: Plugin marketplace governance models need updating to handle AI-generated content and community contributions at scale - Security: Headless AI operation creates new attack vectors that require updated security frameworks for enterprise deployment - Education: Developer education must include AI workflow design and multi-model orchestration as core competencies - Funding: Investment in AI tooling infrastructure should prioritize extensibility and community ecosystem development over feature completeness ### 🧠 Learning Styles Assessment Reading level estimate: 12th grade #### 👀 Visual learners - Present: Live terminal demonstrations, plugin repository exploration, VS Code integration showcase - Improve: Architecture diagrams showing agent coordination, flowcharts for fleet mode decision trees #### 🔊 Aural learners - Present: Live conversation format, verbal explanations of complex workflows - Improve: Audio-only version highlighting key CLI commands and their pronunciation #### 🗣️ Verbal learners - Present: Clear analogies ("fleet of agents"), technical term definitions, conversational explanations - Improve: Glossary of CLI terminology, comparison tables between different AI models #### 🧩 Kinesthetic learners - Present: Live coding demonstrations, hands-on plugin installation, real-world automation examples - Improve: Step-by-step installation checklists, practice exercises for skill creation #### 🧠 Logical learners - Present: Release timeline analysis, systematic feature categorization, clear decision trees for tool selection - Improve: Performance benchmarks, optimization metrics, structured comparison matrices #### 🧑‍🤝‍🧑 Social learners - Present: Community plugin showcase, collaborative development examples - Improve: Discussion prompts about workflow sharing, team adoption strategies #### 🧘‍♀️ Solitary learners - Present: Individual productivity focus, personal workflow optimization - Improve: Self-assessment questions about current tool usage, reflection prompts for workflow improvement #### 🌿 Naturalistic learners - Present: Ecosystem thinking about plugin communities, organic growth patterns - Improve: Analogies to natural systems, evolution metaphors for tool development ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive link text in references - Improve: Alt text for terminal screenshots, ARIA labels for complex demonstrations #### Keyboard navigation - Present: Terminal-focused content naturally supports keyboard users - Improve: Keyboard shortcuts documentation, navigation tips for CLI interfaces --- ## Full Transcript [00:00] Heat. Heat. [00:52] Hello. Hello everybody. Good morning, [00:54] good afternoon, good evening wherever in [00:57] the world you're joining from. Welcome [00:59] to this special special edition of [01:02] coding community. Thank you for being [01:05] here. Welcome, welcome. And for those of [01:08] you who are joining for the first time [01:09] and you're wondering, well, what is this [01:11] even? This is a stream that we do to [01:14] highlight some really fun projects. And [01:16] today's project, I was literally just [01:18] telling our amazing guests that it has [01:21] become my boo. Like it is my single most [01:24] favorite project. Uh it's open source. [01:28] It's by GitHub. It's for developers and [01:30] it really is helping people hone in into [01:33] what it means to code now in this [01:35] agentic era. And you might have heard [01:38] about it before. You might have tried it [01:40] when it was first launched, but I'm here [01:42] to tell you it's not the same. and we're [01:45] going to talk to the PM about it. He's [01:47] also graciously brought some demos and [01:50] some info. So, without further ado, [01:54] please allow me to welcome Ryan Hedge. [01:56] Hello, Ryan. Welcome. [01:58] >> Hello. Hello. I'm so happy to be here. [02:00] >> Me, too. Because last time that we did [02:02] something, I'm going to take this glare [02:04] off of my face. My setup is a bit off [02:06] today, friends. It's been a day. But, [02:09] let me tell you, me and Brian, we [02:10] recorded a GitHub checkout. Uh it's been [02:14] months. I mean it was last year [02:18] maybe [02:18] >> five and a half months from now. Five [02:20] months. Yeah. [02:21] >> Yes. And you know I mean I am I'm not [02:24] going to share that one because but go [02:26] back on your on your own time and and [02:27] and have a listen take a look. But the [02:32] things that you showed me then and the [02:35] way that the CLI is today, [02:37] >> it's I don't know [02:42] aart like it is wild. [02:45] >> It's it's absolutely crazy. There's been [02:47] hundreds of new features that have been [02:48] added. I mean here, do you want to go to [02:50] my screen? Let's let's go look at the uh [02:52] >> Let's do it. [02:53] >> Let's let's go to the uh the repository [02:55] here on GitHub. So I'm going to go to [02:58] the releases. So you see there's been [03:00] 167 releases since um I don't even think [03:04] that includes the pre-releases but you [03:05] know we went all the way back you know [03:08] 925 926 this is when we had our initial [03:11] public preview date there are 17 pages [03:13] of releases and the lengths of some of [03:16] these change logs as well and like how [03:18] many things have been added is is [03:22] insane. Um here actually the easiest [03:25] thing to do I think would just be to go [03:27] to the changel log.mmd. [03:29] >> Yes. [03:29] >> And you can go through and look at look [03:32] some days we ship more than once. Right. [03:35] And this is you know on a daily basis [03:38] all these uh new features and bug fixes [03:41] and uh community suggested feedback that [03:43] we get uh just you know crazy new ways [03:46] of working with agents uh that all just [03:48] go filter into the product each and [03:50] every day. It's it's it's absolutely [03:53] insane. The team's grown so much as well [03:56] since um uh since we first started our [03:58] public preview. Um like you know there's [04:01] there's there's so many designers who [04:03] are helping us make it the thing way [04:05] more ergonomic to use. Um and just [04:08] everyone is so driven, everyone is so [04:11] talented and smart. Um, and like we want [04:15] to make the best uh the best agent best [04:17] agentic harness on the market and I [04:18] think we are uh I think we're succeeding [04:20] at that. [04:21] >> Oh, 100%. First of all, when I say this [04:24] team is cooking, it is no [04:26] overexaggeration. [04:27] The pace I think of obviously the [04:30] community, but the pace in which the [04:31] team has come together to make this an [04:34] incredible experience for everybody. [04:36] Here's a little known fact or maybe it's [04:38] a well-known fact. Um, guess what [04:40] friends? We have to use that. So [04:42] obviously you are cooking with the thing [04:45] that you are building and you want it to [04:47] be a good experience for yourself. So [04:50] I've literally I just mentioned this to [04:52] you before we got on the stream. I'm [04:53] like I don't think I've opened insiders [04:55] like I think the last time I opened it [04:56] was a couple of weeks ago. I was giving [04:57] a presentation and mostly like it made [04:59] more sense to show my file explorer and [05:01] things from there. But I've been leaving [05:04] in Ghosti which is my preferred terminal [05:06] right now because the copilot CLI has [05:09] become my workhorse. Like there is [05:12] nothing that I can't do there that I [05:15] want to do. Like it's amazing. [05:18] >> It's true. I I think if anything like [05:20] you know the rise of terminal based AI [05:21] assistance has reintroduced people to [05:24] just how powerful the terminal is, [05:26] right? And just how you you can do [05:28] anything that you want to do on your [05:30] computer without leaving your terminal. [05:32] Um, and you know, like I' i've seen [05:33] people be like, "Oh man, I wish I wish I [05:35] didn't have to open my editor to, you [05:37] know, like view files or like open my [05:40] file explorer to see files." It's like, [05:42] well, like there are terminal based file [05:44] explorers, you know, like here, go learn [05:46] Neoim and and and and and view your [05:48] files that way. Like it's it's [05:50] reintroducing people to um I think you [05:52] know the most efficient way to uh to use [05:54] your computer and that's always really [05:56] fun. [05:56] >> That is a fantastic way to look at it. [05:58] Honestly, I think I've if anything and [06:00] I'm I'm one of those really like I call [06:03] myself a double a double brain people, a [06:05] spicy brain people and for me to be able [06:07] to just hyperfocus on one terminal like [06:10] it's been just amazing. Amazing. For [06:13] sure. [06:13] >> Amazing. So, when we first launched, I [06:15] can't believe it's only been September. [06:18] This is like what [06:19] >> I know. When we first launch, uh, it [06:22] came out the bag with some real power, [06:24] with some agentic power, with some slash [06:26] commands, but since then, this thing has [06:30] exploded. Can you tell me a little bit [06:32] about what's new new? I mean, just [06:35] yesterday I was I I thought the most [06:37] recent theme was fleet and delegate and [06:39] then I saw Evan post a tweet and I'm [06:41] like, "Oh my, I can't even keep up." So, [06:43] please spin. Let us know what's new. [06:46] >> Oh, boy. Yeah. So, I mean, I want to [06:48] start, you know, obviously we could [06:49] delve into just the things that have um [06:50] that have that have filtered into the [06:52] product like in the past week because [06:53] there's cool things there, but I mean, [06:54] let's take a step back uh and maybe tell [06:56] the people just, you know, [06:58] >> what we're talking about, what's [07:00] happened over the course of the past few [07:02] months um here with the Copilot CLI. [07:05] Yeah. [07:05] >> So, you know, obviously you [07:10] >> you can talk to it, it can edit files on [07:13] your computer, it can run any command [07:14] that's on your computer. Um, but because [07:17] it is in the terminal, you can run it [07:20] headlessly. You can include it in [07:22] automation, right? So, you can have a [07:24] gentic assistance uh, you know, just [07:26] kind of in a script that's on your [07:27] computer. Um, you can extend it lots of [07:30] different ways, and I'll show you a few [07:32] of the ways that you can do that, um, a [07:34] little bit later on. Uh and you know you [07:37] can really customize it to your use case [07:40] uh with hooks with um uh you know with [07:44] custom instructions uh with MCP servers [07:47] uh and you can you can run it anywhere [07:49] um that you can you know log into GitHub [07:51] or you can just mint a personal access [07:52] token and and and put it wherever you [07:54] want. Um, you know, I've seen everybody [07:56] go uh crazy over, you know, OpenClaw and [07:59] the likes of these kinds of personal [08:00] assistants. Like you can without very [08:04] much effort um you know, use the Copilot [08:07] CLI to be something that powers that for [08:10] you and then just kind of create some [08:11] wrappers and some customizations around [08:13] that. Uh and I've done something like [08:14] that uh for myself as well. And it's a [08:16] it's a really great way to kind of I [08:18] think organize organize your your your [08:21] life, you know, like personally, [08:23] professionally, or or what have you. [08:25] >> Yes. All of these beautiful core [08:27] features that are the things that come [08:29] with the harness, right? Like it's is is [08:31] ready to work. And I know you're going [08:33] to talk about what people have access to [08:35] because I think one of the things that's [08:36] we don't talk about enough is how you [08:38] can literally work with so many models [08:40] using the GitHub copilot CLI. So you are [08:43] not tied to a single model. You get to [08:45] choose [08:47] what it is that you want to put on to do [08:49] the work that you like. We all have our [08:51] preferences. They're all like kids. I [08:52] love them all, but I have my favorites. [08:54] >> It's true. It's true. Yeah. You know, [08:56] like Yeah. And you know, some people are [08:57] like, "Oh, I like how, you know, Codex [08:59] 5.3 is a little bit more tur and to the [09:01] point." Or some people don't like that [09:03] and they like how Opus 4.6 is a lot more [09:05] kind of free form and we'll go off on a, [09:07] you know, several paragraph um diet [09:10] tribe about whatever you want it to. Uh [09:11] but I think really the thing is um is [09:13] that choice, right? and is being able to [09:15] go through, you know, we've all seen the [09:17] um uh we've all seen the little [09:18] flowchart meme of, you know, every three [09:21] months whoever has the leading model [09:23] changes. Uh and I think um being able to [09:28] when a new model ships know that you [09:30] have one subscription that's going to [09:32] get you access to all of them, that's [09:34] huge. Um but even more than that, I [09:36] think being able to dispatch [09:39] multiple models at the same task is a [09:42] big deal, right? What I'll find myself [09:44] doing is uh using plan mode to come up [09:47] with a plan for a feature that I'd like [09:49] to work on. And then before I start [09:51] implementing the plan, I will say here [09:55] spin up parallel agents using Opus, [09:58] using codeex, and using Gemini and have [10:00] each of them review the plan and give [10:02] feedback on it as a little pro tip uh [10:04] for anybody. And then, you know, then [10:06] it'll come through and it'll say, "Okay, [10:08] here's where the three models agreed. [10:10] Here's where they disagreed. would you [10:12] like to implement some of the [10:13] recommendations of each of them? And [10:14] then you have this like uber strong [10:17] ready to go plan. Um and then it can go [10:19] and it can implement it. And then after [10:21] it's implemented it, then you can go [10:23] into review mode with each of those [10:25] agents in parallel as well and say, hey, [10:27] what does Opus uh you know um codeex and [10:30] Gemini think of how this was [10:33] implemented? And you know, do we want to [10:35] make any changes to the implementation [10:37] based on based on their recommendations? [10:39] Model choice is huge and um you know [10:42] obviously you can get model choice by [10:44] having a cloud code subscription, having [10:46] a codec subscription, having a Gemini [10:48] subscription and uh ping ponging between [10:50] each of the apps. But I think it's great [10:52] to be able to do all of that in one [10:55] harness without moving. Um that's a that [10:58] that that's a big strength of of Copilot [11:01] for me. [11:01] >> Absolutely. One hardness, one memory. [11:04] We're going to talk about memory, too. [11:05] We got to talk about all the things that [11:07] I and there's more to come but I just [11:09] want folks to like be cognizant of that [11:11] because I think [11:12] >> it's tough like and honestly and [11:14] especially you brought up open like so [11:16] many people are saying like oh I've had [11:18] I stopped paying for like five different [11:20] subscriptions because I just buy my own [11:21] thing. I've been doing a lot of like [11:23] this little macro tools and like Mac up [11:25] things that are like literally solving [11:27] my own problems. The ones that are not [11:29] terrible, I've open source. So, check [11:30] out my profile. But for the most part, I [11:34] love being able and empowered to do [11:35] that. And to have that sort of like [11:38] sense that I don't need to go pay 10 [11:40] different providers. I have the best [11:42] agentic harness giving me the best [11:45] possible models. [11:47] What else I want? [11:48] >> Exactly. Exactly. And knowing that when [11:50] the next best model comes out, like [11:52] you're going to have access to it, uh, [11:53] you know, like nearly instantaneously. [11:55] We've gotten real fast. That's [11:56] >> I know. It's impressive when you see it. [11:59] No, no, no, no sooner I've read the [12:01] tweet that I already know is about to [12:03] pop up as an option for me. [12:05] >> It's true. It's true. It's true. I [12:06] really love to see that. Um, but yeah, [12:09] you know, and you can use it [12:11] interactively in your terminal and [12:12] that's obviously great and that's the [12:14] way that um I think I personally [12:16] interact with it the most and I think a [12:17] lot of folks do as well. Um, but you can [12:20] also run it headlessly. uh you know so [12:22] we will uh utilize this to um help [12:25] triage issues that come in um in our [12:28] various repos right we'll have copilot [12:30] being run headlessly in um in a GitHub [12:32] action that then you know triggers some [12:35] kind of uh it'll apply labels to it'll [12:37] be like oh I think this is a duplicate [12:39] of x y and z right um uh or even you [12:42] know if if it falls into a certain class [12:44] of of issue maybe could just [12:45] automatically assign co-pilot uh the [12:48] cloud coding agent to go and try and [12:49] solve it right Um, and then being able [12:51] to um to to to run it headlessly like [12:54] this is is huge for um for interact for [12:58] non-interactive uh workflows. Um, even [13:01] more than that though, um, is this [13:03] notion of, and we don't have amazing [13:05] docs on this yet because it's kind of a [13:07] little bit more experimental, but um, [13:09] uh, using ACP and all of the, um, all [13:13] the different JSON events that you can, [13:15] um, use to, uh, to drive that, you can [13:18] run Copilot in this server mode and then [13:20] you can have other clients go and drive [13:22] your CLI session. Um and so you could go [13:25] and and and you know this could be a way [13:27] that you could implement you know an [13:29] open claw like this is the other thing [13:30] there's so many different ways of going [13:32] about solving these problems. Uh and [13:34] there's lots of different ways that you [13:35] can go and um uh um uh and use the tool. [13:39] Uh and you one thing you can do is you [13:41] can do copilot-server [13:42] and now um uh copilot can communicate [13:45] with um you know an editor with another [13:48] application um or some other agent [13:50] that's going to be driving the show. Um, [13:53] and this is super coming soon, but you [13:54] know, we want the idea to have um [13:57] multiple clients drive the same session [14:00] together. [14:02] >> I love that. This is sort of like so you [14:04] could have [14:05] >> not only many many many sessions running [14:08] on parallel too, but then multiple oh [14:10] multiple like instances in a way. This [14:12] is really cool. [14:14] >> Yeah. Yeah. I think flexibility is the [14:16] name of the game, right? And making sure [14:17] that no matter how you want to work with [14:18] an agent, we have a solution for you to [14:21] plug it in there. Um, you know, it's [14:22] great if you are doing all of your work [14:24] on github.com and then, you know, all [14:26] you need is the the CLI on your local [14:28] computer and then you push it up and you [14:29] got, you know, the cloud agent um, up on [14:31] GitHub. But, uh, there's so many [14:33] different ways of working and there's so [14:35] many different ways of embedding agents [14:37] into your daily work. Um, and we want to [14:39] give you all of that, all that [14:40] flexibility. [14:41] >> I love that. And honestly, Ryan, and [14:42] shout out, massive shout out to the team [14:44] and GitHub and you because I love the [14:46] way that we're playing nice with the [14:47] ecosystem as well, like adding ACP as a [14:50] feature. Like that was a community [14:51] request. Oh, yeah. [14:52] >> Uh because people wanted to, you know, [14:54] they wanted to use Zed. They wanted to [14:56] like bring it into Nub, you know, now [14:58] it's shipped like developers and [14:59] adopting it and I love that we're doing [15:01] that. So, [15:01] >> yeah. Yeah. And we got to we got to [15:03] collaborate uh with our friends at Zed [15:04] as well to bring this uh to to to life. [15:06] And I think it's um it's it's good to [15:08] make um to make this space feel less [15:12] walled gardeny. Uh you know, I think [15:14] there's a there's definitely always um [15:16] you know, the pressure to kind of do [15:18] that. And I think um what something I [15:20] really like about our ecosystem is that [15:22] we don't have these walled gardens. We [15:24] do let you kind of plug and play and and [15:26] and choose how you use um uh the [15:28] different the different products. [15:30] >> I love it. I'm team use what makes you [15:32] happy. [15:32] >> Oh yeah. Okay. What else can you show [15:34] us? Well, look, I and I think, you know, [15:37] it's about way more than just coding as [15:38] well. [15:39] >> Yes. [15:39] >> Right. So, yes, you can generate code [15:42] and that's the classic use case for [15:43] these things. uh you know but I think um [15:48] something that I've been using um this [15:50] tool a lot especially in the world of [15:52] connecting with um you know things like [15:55] Slack or Teams or your email or your [15:58] GitHub issues is being able to automate [16:01] a lot of like knowledge work kinds of [16:04] tasks. Um so I'll give you a little [16:07] example. Um we have a um we have a [16:11] spreadsheet that's right that's in [16:13] Google Sheets. We're in the process of [16:14] migrating it into Excel. That's a longer [16:16] story. Um, but you know, we kind of just [16:18] capture a bunch of metrics about our [16:19] product every single week. And those [16:21] metrics come from um, uh, Azure data [16:25] explorer in acustoto. Uh, query is what [16:27] is what we use to generate those. And, [16:30] um, you know, we need to report on those [16:32] metrics once a week. [16:35] Usually, this is just a very annoying [16:36] manual step that you have to do because [16:38] I have to go to the to the dashboard. I [16:40] have to make sure I'm selecting the [16:42] right rows of the of the of the um of [16:45] the graph. Maybe it's a graph, but I [16:46] need to get it into a table so I can sum [16:48] things up or whatever. And you know, [16:50] it's a very manual process. Eats up [16:53] time. It's annoying and it stops me from [16:55] you know, engaging with a lot more um [16:57] you know, interesting problems to solve [16:59] at work. uh the copilot CLI [17:03] uh can use the um uh the Azure MCP [17:07] server to run the custo query for me in [17:11] the format that it needs, grab that [17:13] number and then uh drive a session in um [17:17] in in in in Google Docs to go and paste [17:19] that number into the right cell of the [17:21] spreadsheet. This is something that like [17:23] anyone can just build and if you don't [17:25] know how to build it, you can talk to [17:26] co-pilot to create a system that lets [17:29] you do these things. [17:30] >> Wait, wait, wait, wait, wait. So, not [17:32] only because anyone can do a copy and [17:34] paste, but this is you're actually the [17:36] agent itself, like the CLI itself is [17:38] actually [17:39] >> going into the correct cell [17:41] >> and doing what it needs to extract the [17:43] data and then it's going to present it [17:44] to you the way that you ask. [17:46] >> Yeah. Yeah. And it's it's much easier to [17:48] do that now with um uh you know with uh [17:52] with some other tools that um that that [17:54] Microsoft made. The work IQ plugin for [17:56] the copilot CLI. Has anyone have have [17:58] you shown that off on uh on one of your [17:59] streams? [18:00] >> We haven't a lot. No, I've been playing [18:02] with the SDK like the office one like [18:06] making uh powerpoints from nothing. But [18:09] I know work IQ is even it takes that [18:11] even to a next more deeper level because [18:14] then you have the entire view of [18:16] everything like literally everything [18:18] that you have your emails. Tell us a [18:21] little bit about it. [18:22] >> Yeah. Yeah. Um so work IQ is a system [18:27] that allows you to query anything that's [18:29] happening in your Microsoft 365 [18:31] ecosystem with natural language. So you [18:33] could say uh you know hey are there any [18:35] emails that are waiting on responses for [18:37] me that I haven't gotten to yet? or what [18:39] meetings do I have going on next week? [18:41] Uh or uh what documents have I been [18:44] pinged in? Um and these kinds of things. [18:46] And you ask it in natural language and [18:48] it spits back uh you know like real [18:51] responses um you know about your about [18:54] your work life along with links to where [18:57] you can go and action on those. Um and [18:59] that's huge in and of itself. So, you [19:01] know what? If that was made available to [19:04] you via an MCP server and then you could [19:07] use it in the C-pilot CLI and ask, you [19:11] know, like what are my upcoming meetings [19:12] this week and then, you know, you just [19:15] have your readout right there in the CLI [19:16] and you don't you didn't need to go and [19:18] open Outlook to uh to try it out. Um, I [19:22] I really should set up a um like a demo [19:25] uh uh Azure tenant so I can actually [19:28] demo this publicly without risking uh [19:30] sharing any confidential information. Um [19:33] >> but uh trust me, it's really cool. Uh [19:36] and all you have to do if um uh if you [19:40] if your company uh uses Microsoft 365 uh [19:44] and you have a GitHub Copilot [19:45] subscription, download the CLI and run [19:48] these two commands. Um I think he just [19:50] put the uh the link to the docs in chat. [19:53] Yep. Uh so open up Copilot, run these [19:55] two commands um to add uh this [19:58] marketplace and then install the work IQ [20:00] plugin and then you log in and then you [20:04] can just start asking it about all of [20:05] your work uh information. It is it's [20:07] it's it's honestly very magical. [20:10] >> I love this. Honestly, this is the way. [20:12] This is the way. I know there's a lot of [20:14] things and a lot of us have our own [20:15] workflows before this for sure where we [20:18] have different tasks running in [20:20] different schedules like you have your [20:22] email checker you have your calendar [20:24] kind of create an agenda for me but you [20:27] don't need to create authentication [20:29] rules for any of those things or [20:31] actually create separate endpoints or [20:33] even throw APIs in there because work IQ [20:36] does it for you with MCP server [20:37] integration and bringing in copilot CLI [20:40] you have now a full view of everything [20:42] >> it's [20:44] That's cool. That's cool. [20:46] >> Let's take a question from Ben. And [20:47] thank you so much, Ben, for being here. [20:49] It sounds similar to 365 copilot. I [20:52] know. Listen, naming things is hard and [20:53] we get confused. Okay. But is it a [20:56] different thing? [20:58] >> So, uh, Work IQ is a um I Okay, so it's [21:04] essentially an interface for the [21:06] Microsoft 365 copilot. And so instead of [21:08] needing to go on the web and talk to uh [21:11] Microsoft 365 copilot, you know, uh on [21:13] whatever their website is, quite [21:15] honestly, I've never uh interacted with [21:16] it on their website before. The Work IQ [21:18] CLI is the only way that I go about [21:20] doing it. Um but uh you know, this [21:22] essentially liberates that from just [21:24] being um uh in the existing places where [21:28] Microsoft makes it available for you. [21:30] And now you can just go into the CLI and [21:33] access it anywhere. And because it's in [21:34] the CLI, it literally can be accessed [21:37] anywhere. Like literally anywhere that [21:40] you can get a terminal running um or you [21:43] know just like you know spin up a [21:44] co-pilot process which is basically [21:47] anywhere. [21:47] >> I love that. And listen this is all [21:49] powered by the fact that they have an [21:50] MCP. This is one of the MCPs that you [21:53] can 100% like point your CLI your [21:56] copilot coordination to. Like that's [21:58] cool. Like that there's a lot of [22:00] interoperability. Well, I can't say that [22:02] word. uh and a lot of uh plugins. Oh, [22:05] let's see about awesome copilot. And I [22:07] do want to take Evo. I'm coming to your [22:08] question because that's going to be a [22:09] deeper conversation about creating [22:11] custom agents and I want to make sure [22:12] that we actually we might even demo that [22:14] Ryan if you're up for it. [22:15] >> I I we absolutely can. Yeah. Yeah. Um [22:17] and but you know I if you if making your [22:21] own skills or custom agents uh seems a [22:24] little intimidating to you, which I can [22:26] understand. Um it's not as scary as it [22:28] sounds. So we'll, you know, we will get [22:30] to that. But um there is a whole [22:33] community of people out there who have [22:35] already created a bunch of really [22:37] amazing things um that you can start [22:40] using right now uh to turbocharge your [22:42] usage of the CLI or any GitHub copilot [22:45] uh surface to be uh to be quite frank. [22:47] Um, but if we go into and I really love [22:50] plugins because plugins, [22:52] okay, [22:54] >> hand up if you have wondered when it's [22:56] the best time to use an MCP server or a [22:59] custom agent or a skill. [23:01] >> It's confusing, right? And it's like, [23:03] you know, like and you don't know, you [23:05] know, when people are making your tools [23:06] available, if you have to add something [23:07] to an MCP config, if you got to install [23:09] some custom agents to your repo or [23:11] skills, like it's just a lot to it's a [23:13] lot to remember. It's a lot to keep [23:14] track of and they all get a little a [23:17] little muddled in the water. [23:18] >> Um, plugins [23:21] solve that in in a way. Um, because [23:25] plugins allow you to essentially um here [23:27] let's let's just pick one of these. Um, [23:31] what is what is noob mode? I don't think [23:34] I've seen noob mode before. Uh but it [23:37] essentially allows you to bring together [23:40] um all these different extensibility [23:41] concepts and make it so all people need [23:44] to do to use it is run one install [23:46] command. Uh so that work IQ uh plugin [23:49] plugin you know and all someone needs to [23:51] do is add your marketplace and then [23:53] install the plugin and then all of a [23:55] sudden they just have access to the [23:57] tools that you made available. Um, and [23:59] so this noob mode plugin, uh, which you [24:01] would just install by, um, installing [24:03] noob mode at awesome co copilot, um, it [24:08] gives you a, uh, a command that explains [24:13] everything in plain English. So that's [24:15] very nice. If you're if you happen to be [24:18] new to uh to programming, if you know [24:20] this is your first on-ramp into that, um [24:22] it can go through and uh and and and and [24:26] translate things in a way that a little [24:27] bit easier to understand. [24:30] >> That's pretty cool. [24:31] >> Look at all of those. And this is a also [24:33] an open source project, friends, where [24:34] people are contributing, but this is [24:36] curated by part of the both Microsoft [24:38] and GitHub teams, right? So, it's going [24:41] to be nice even though it's it's it's [24:44] open and folks can submit like there is [24:46] a ton. I mean, how many skills I need to [24:48] look it up like [24:50] >> how many skills are in here? [24:51] >> There is like a ton, right? The plugins, [24:53] just the plugins itself. There is [24:55] >> Yeah. Well, yeah. Remember that all the [24:57] plugins might include skills, but if we [24:58] just go to the skills folder here, [25:01] >> that's a lot. [25:02] >> Oh, boy. There's a lot of these. [25:04] >> Yeah, that is a lot. And and this follow [25:08] the same convention that skills are [25:11] written on. So you can use them [25:14] anywhere. [25:14] >> It's true. It's true. And and remember [25:17] when I said it it's not that scary to um [25:21] to write a um to write a skill. All you [25:25] have to do to write a skill is set a [25:29] name for it, give a description for it. [25:31] This way, the agent knows when, you [25:33] know, in order to uh when it's a good [25:36] time to invoke this skill. And then you [25:38] just explain what you want the agent to [25:41] do. And so, in this noob mode one, um, [25:43] which again, I I've never seen noob mode [25:45] until a few minutes ago when we clicked [25:47] on this. Um, but but all the author is [25:49] saying is, "All right, Copilot, I want [25:51] you to activate noob mode to make the [25:53] Copilot CLI speak plain English. It's [25:56] designed for non-technical [25:57] professionals." And it gives you [25:58] examples of uh of of of who some of [26:01] those might be who use the copilot CLI [26:03] but don't have a software engineering [26:04] background. When new mode is active, [26:06] translate every permission request, [26:08] error message, and technical output into [26:10] clear jargon-free language. And then it [26:13] gives some ex some um uh features of [26:15] what it does. So you just literally [26:18] describe what you want it to do and then [26:21] it knows how to do it. Um like what I've [26:24] done is I've also taken um uh for some [26:26] teams that I work with like uh tutorials [26:29] um on how to use certain pieces of [26:31] software and tutorials on how you know [26:33] what our standard operating procedures [26:34] are for things. Just put them in a skill [26:37] file, slap a name and description on [26:39] there and see what happens. And you you [26:40] might be surprised at how um at how good [26:43] of a of a of a job um you know co-pilot [26:46] can then just do in um in following [26:48] that. [26:49] >> Oh, I love that. I love that. That's and [26:52] honestly like if you're looking for [26:54] something that you think might be it'd [26:56] be a good skill to have like go in there [26:58] and do a search there's been some really [27:00] amazing skills that have been [27:01] contributed. I mean just for about [27:04] anything you can think of there is a [27:06] skill. Um and again it's all about the [27:09] way that they're sort of laying out the [27:10] instructions for the models. I just I [27:12] was just listening to some a podcast a [27:14] little bit ago and person in the podcast [27:16] mentioned how it's like we try to give [27:18] constraints to the models when you [27:20] really need to just point to them and [27:22] let them go especially some of the you [27:24] know the the higher reasoning models and [27:26] I thinking models like they can figure [27:28] it out. You just got to kind of be clear [27:30] on your instructions and setting up the [27:32] skills is one way that you can really [27:33] make that sort of easier for them. So [27:36] very cool. uh when is skills are too [27:40] much is that kind of a thing. I mean you [27:43] can have any amount of skills and that's [27:44] another thing too like I feel like this [27:46] does this apply like dynamic tool [27:48] calling like where it will actually use [27:49] the skill when you when you ask it to [27:52] Yes. [27:52] >> on the CLI. Yeah. So you can have [27:55] however many skills um installed. [27:58] >> Yeah. Yeah. Now um you know obviously [28:01] there is such a thing as having um it's [28:05] more of a big deal with MCP servers than [28:06] skills but you are using up part of the [28:09] models context window when you install [28:11] these tools. Um and actually if you go [28:13] into the uh [28:15] >> if you go into the copilot CLI and then [28:18] if you run context [28:20] um well after my um I also want to put [28:24] streamer mode on for fun because I'm I'm [28:26] in streamer mode. [28:27] >> Shout out for the team for doing that. [28:29] Thank you. [28:29] >> That was great. Yeah, that was great. [28:30] >> Uh so if I do slashcontext in the CLI, [28:33] um it will tell me how much of um my [28:36] context is currently used up by um uh by [28:40] tools that I have defined. Um and so you [28:42] know there there is such a thing as [28:43] having too much because then you have [28:45] less space in your context or your [28:47] actual session. Um and then you know [28:49] whatever it'll start going to compaction [28:51] a little bit more quickly. Um but skills [28:55] are a lot better than MCP servers for [28:57] that. Um, and I would also encourage you [29:01] if you have a lot of MCP servers and you [29:02] notice that, hey, my agent, uh, you [29:04] know, it starts, uh, it starts getting [29:05] into compaction a lot earlier than I'd [29:07] like, um, look into whether there are [29:09] any CLI equivalents for your MCP [29:11] servers. Um, there's even some tools out [29:13] there that will help you convert an MCP [29:15] server to a CLI and then you can just [29:18] provide a skill of here's a [29:20] documentation on how to use the CLI. Um, [29:22] and that's a lot um a lot less tokens [29:26] that you'll end up being using in uh in [29:28] your context there. [29:30] >> That is a massive tip. I think like I'm [29:33] super hype and bullish on MCP, but I [29:35] mean like there are some things that you [29:37] don't need the full MCP server to do [29:39] what you need to do and skills might be [29:40] the perfect solution for that. [29:43] >> It's true. Yeah. Yeah. Um, someone asked [29:45] in chat, uh, if plugins can include [29:47] custom instructions or are they just [29:49] limited to agents and skills? [29:52] >> Yes. Thank you for the question. [29:53] >> Yeah. Yeah. So, I'm going to go to, uh, [29:56] create a plugin here. Um, so you can [29:59] have agents, skills, and hooks. Those [30:02] are the three things that you can put [30:03] into your um, uh, into your plugins. [30:06] Now, custom instructions. Custom [30:09] instructions are good for uh putting in [30:12] your in your repository for example. So [30:14] that way GitHub knows how to better uh [30:17] do work in that uh repository and in [30:20] that codebase. Um or even at a user [30:22] level like say you want Copilot to talk [30:24] to you in a certain language or you know [30:26] in a certain response style um it's good [30:29] to have custom instructions uh at the [30:31] user level for that. [30:33] in the in the case of a plug-in, um, [30:37] skills are really what you want and not [30:40] custom instructions because you don't [30:42] want it to be acting that way all of the [30:44] time when it has the plug-in installed, [30:46] right? Because someone might have your [30:47] plug-in installed and be doing a whole [30:49] number of different things. They could [30:50] be doing, you know, software [30:52] development. They could also be doing [30:53] creative writing, right? Um, or they [30:55] could be doing, you know, systems [30:56] administration. and you probably don't [30:58] want to have custom instructions uh in [31:00] there kind of affecting how the agent [31:02] responds in those situations. Instead, [31:04] you have a an agent that then allows you [31:07] to um uh to take on that persona. So, [31:12] anything that you would put in custom [31:13] instructions in the plug-in, you might [31:15] want to put it in a custom agent [31:17] definition or in a skill file uh [31:20] instead. [31:22] >> Perfect. Thank you so much for that and [31:24] thank you for the question. Let me go up [31:26] a bit because this is also another good [31:28] question. Thank you again from empty [31:31] freight tag. Thank you. Uh can they only [31:33] be installed on the CLI or are they [31:35] coming to VS Code? [31:37] >> So um I was actually just told today [31:39] that VS Code has shipped plugins. Um I [31:42] don't know how it works. I haven't tried [31:43] it yet. But uh they do have plugins. Um [31:46] here we even VS Code Copilot plugins. [31:51] Let's see. Is it documented yet? Because [31:53] it might not be. It might be. I mean [31:55] just yeah brand brand new there was just [31:57] a release of of code like maybe [32:00] yesterday. [32:01] >> Yeah VS Code Insiders got jealous of how [32:04] fast we were shipping and they've [32:05] started shipping very fast too. [32:07] >> I love it. [32:09] >> Uh but yes uh so they are they are [32:11] available in VS Code. I'm sure it's a [32:13] matter of um uh you know they'll be [32:16] adding um more functionality as time [32:18] goes on and everything. um uh plugins [32:20] are in this weird wild west space where [32:22] there isn't uh an official spec for them [32:25] yet. Uh there isn't really a standard [32:26] and so I know that um certain harnesses [32:28] might have slightly different um you [32:30] know differences between uh between how [32:32] plugins work and everything but um uh [32:35] you know unlike skills skills do have a [32:37] standard and um it's agentskills.io.io [32:41] io uh io. Yeah. Um so I would invite [32:46] anybody who's interested in uh writing [32:48] skills and want to learn more about tips [32:50] for how to do that and the different [32:52] kinds of references you can give them uh [32:54] go check out agentskills.io and they [32:57] have the whole the whole spec there. Uh [32:59] or even better take this um go into [33:03] copilot and say tell me more about how [33:06] to make good skills and just paste in a [33:11] link uh and then it'll just go and it'll [33:13] tell you about you know it's it's it's [33:15] it's always great to see like that. [33:17] >> Look at that. And I honestly this is [33:19] like this has been my view and it's and [33:21] it did take a bit of a getting used to [33:23] to be honest like it did but now I'm [33:25] like I don't want to work any other way [33:27] like [33:27] >> it's true. No, it it really is true. It [33:29] really is true. I think the the ability [33:31] to [33:32] >> talk to your tool and learn how best to [33:35] use it is a big deal. Um, and we've [33:40] actually made it way easier to do that. [33:43] Um, this this might I might become I [33:46] might be very embarrassed by doing this, [33:47] but we're just going to do it. [33:49] >> Just yolo it. [33:51] >> I'm going to go into co-pilot. I'm going [33:53] to do slash chronicle. Have you played [33:55] with slash chronicle yet? just saw it [33:57] today and I have not because I'm kind of [34:00] like scared of what I might find. [34:03] >> It is true. It Yes, I I totally [34:07] understand that. Uh but if I do SLC [34:09] chronicle tips, what Copilot's going to [34:12] do is it has indexed all of my [34:15] conversations that I've had with it. Um [34:17] and it's going to go and take a look in [34:20] all of those conversation transcripts. [34:22] It's going to check the documentation [34:24] for GitHub Copilot uh CLI and look at [34:27] how I use it, the type, you know, the [34:29] way that I talk to it, the the the [34:32] skills that I use, um the uh the types [34:36] of tasks that I that that that I go [34:38] about doing, and it's going to tell me [34:40] how, you know, well I am utilizing the [34:42] product. It's going to tell me, you [34:43] know, here are some things that um you [34:45] know, you might you might want to [34:47] change. You know, I've seen that you [34:48] you've done the same task over and over [34:50] and over again. Why not make a skill for [34:52] it, right? Uh, you know, I see that um [34:56] there's a way that you that you describe [34:58] this to the agent uh leads to lots of [35:01] back and forth to better understand what [35:02] you're talking about. Here's a way that [35:04] you might want to, you know, condense [35:06] that into some kind of custom [35:07] instruction uh to make that be uh better [35:09] used. Um there's there's lots of ways [35:11] that we can improve. And yes, it could [35:13] be difficult to get some of that uh [35:14] constructive feedback and be like, "Oh, [35:16] yeah, you know, don't I feel silly?" [35:18] >> Yeah. No, but this is really [35:19] interesting. I honestly I haven't done [35:21] it. I think I don't know when did I [35:22] check it because I literally I remember [35:24] just seeing it and putting mental no you [35:25] need to do that. Yeah. [35:26] >> Um I've done that with copilot before [35:30] where I asked it after we've been [35:31] working on something for a very long [35:32] time and I asked it okay like how could [35:34] I prompted you better so we got here [35:36] sooner. It's been kind of like that [35:39] mental conversation but now I don't even [35:40] have to think about it because I can use [35:42] the chronicle hack and then it'll go. [35:44] >> It's true. It's true. Yeah. Yeah. And of [35:47] course, I also um nuke my my config all [35:50] the time as I'm testing different [35:51] things. So, this isn't a full picture, [35:53] but at least what's currently in my [35:54] config file. Um you know, apparently 53% [35:57] of my sessions are single turn. Um and I [36:00] just ask questions and I'm done. I'm not [36:01] making great use of our infinite [36:02] sessions concept these time uh right [36:04] now, but you know, um you know, I'm I'm [36:07] triaging user reports. I'm debugging uh [36:09] custoto queries. I'm drafting issues in [36:11] PRs. Um and it's it's looking into look [36:14] it says you're already using file [36:15] mentions fleet mode. We haven't talked [36:17] about fleet mode yet. We can get into [36:19] that. [36:19] >> Yes, please. We have to [36:20] >> uh you know custom skills work IQ. So [36:22] these tips are going to skip the basics. [36:24] Um so this is a brand new feature [36:27] research. It essentially is a deep [36:28] research uh command workflow for the [36:30] copilot CLI and it's noticing that I'm [36:33] not really using it. Uh to be fair most [36:35] of the sessions in its index are from [36:37] before research existed. Uh so that's [36:40] why um but you know like yeah it would [36:42] be great to do uh to use that to do um [36:44] uh some some deep research agents to [36:46] look at um things across GitHub. Uh fair [36:49] enough. Um I'm regularly pasting [36:52] multiundline queries and uh really long [36:55] prompts. I should use control G to open [36:58] long prompts in my editor. Did you know [37:00] that you could do that? [37:00] >> I did not know that I could do that. [37:03] >> Yeah. So if I do control G, here you go. [37:05] It's opened it up in Vim, which is uh my [37:07] default editor. um for here. Now I can [37:10] add to it. Hello. This is cool. [37:14] >> This is cool. [37:16] >> And then if I close Vim, now the prompt [37:19] is here. [37:19] >> Ryan, you're brave. You exited them and [37:21] you showed us your chronicle all on the [37:23] same screen. [37:24] >> It's true. It's true. It's true. You're [37:26] right. Uh first try. Um, but yeah, you [37:29] know, like no matter what your editor of [37:32] choice is, if you would rather edit your [37:33] prompts in there, you have all your key [37:34] bindings all set up just the way you [37:36] like them, you have all your extensions, [37:37] you have your autocomplete, [37:39] you can do that now. And it's a lot [37:41] easier to um uh to edit all of your uh [37:44] your crazy prompts. And so, yes, maybe [37:46] when I'm pasting my multiundline custom, [37:49] uh I should be doing that instead. Um [37:52] that's definitely that's definitely [37:53] fair. [37:54] >> That's a heck of a tip. And let me tell [37:56] you something. This is a way that you're [37:57] actually getting feedback to improve the [37:59] way that it's not just for the model. [38:01] Like for the agent, it's for you. Like [38:03] >> it's true. Yeah. [38:04] >> For you. Like I love this conversation. [38:07] Okay. So now new new flag to try. Uh [38:10] remind us what it was. It was Chronicle [38:13] >> the Chronicle Tips is what that was. [38:15] Yes. Yeah. Yeah. And we also have um uh [38:18] Okay. Yeah. Okay. We're going through [38:20] it's this is still doing some deep [38:22] research into skills it seems. Um, oh, [38:25] and then I canceled it. Uh, but okay. [38:27] Well, whatever. We'll go back to that [38:28] later. [38:28] >> You You have them all running on like [38:30] super high context, like you have it [38:31] going like in turboation, [38:34] triple diamond mode. So, it was going to [38:36] take a minute. [38:37] >> Yeah. Yeah. Yeah. [38:38] >> It's going to take a minute. [38:39] >> Uh, there's also Chronicle improve and [38:41] Chronicle Standup. Um, so there is [38:44] different things that um the Chronicle [38:46] Standups, it's pretty simple. It [38:48] basically just lets you um uh it'll look [38:50] at some of the sessions that you've had [38:52] uh going on. maybe you didn't resolve. [38:54] Um so that way in the morning you can [38:56] run chronicle standup. Um and then have [38:59] an understanding of what you should be [39:00] doing. You know, it's a standup, right? [39:02] Um so what I what I find myself doing is [39:04] in the morning I'll do chronicle standup [39:06] to remind me about uh you know where [39:07] I've left off and what I ought to be [39:09] doing right now. Um and then in the [39:11] afternoon or when I'm signing off I can [39:13] do uh Chronicle tips uh to kind of see [39:15] like all right the sessions that I've [39:18] done today uh is there anything that I [39:20] can [39:24] Oh, hang on, Ryan. I think we lost your [39:25] sound. [39:26] >> I muted I muted my microphone uh on um [39:30] uh yeah, that was interesting. Um but um [39:34] what um yes, then I use Chronicle uh [39:38] tips uh to generate tips on how I've [39:41] been using Copilot when I'm kind of like [39:42] ramping down and I have some time and [39:44] they can let me uh go and improve some [39:46] of my uh some of my files. Um and then [39:49] here uh crop copilot sorry chronicle [39:51] improve um that is also then looking [39:54] into um some of the uh the custom [39:58] instructions that it has some of the um [39:59] the other customizations that we've made [40:01] uh and looking for ways that we can [40:02] actually just improve the way that we [40:04] are uh prompting um copilot. Um I [40:07] haven't run this one on on on this batch [40:10] of sessions so I I have honestly no clue [40:11] what we're going to find when uh when it [40:13] finishes. [40:15] >> This is very interesting. It's like [40:16] copiler has become sort of like your [40:18] your mentor manager type b too. Like I [40:23] like that. Yeah. [40:23] >> Seriously, you know, it's like it's like [40:25] like having a little like, you know, [40:26] coach uh with you at all times and [40:28] that's awesome, you know, like it's [40:30] helping you get be able to use the tool [40:32] better and then also just be better at [40:33] the uh the tasks that you're going out [40:35] to do and like I think that's super [40:37] super cool. [40:38] >> While it's cooking, let's take this [40:39] question. Mr. Kale, this is a good [40:41] question. Yeah. [40:42] >> So you have Copilot and VS Code but not [40:44] Copilot CLI. And this is something I [40:46] want to clarify for everyone who's [40:47] watching. You get Copilot CLI too. [40:50] >> Yes, you do. [40:51] >> Everybody gets Copilot CLI. Um do you [40:54] need it? Yeah, if you want to install [40:56] it, install it. And you can you can of [40:58] course open a terminal in VS Code and [41:00] use it there if you want to. There is [41:02] nothing stopping you. But uh yes, and [41:05] it's available to everyone. So [41:08] >> yeah. Yeah, I know that was something [41:10] that the team it was important for the [41:11] team that that was like offer as not [41:14] just as part of the subscription and so [41:17] everybody go download install the CLI [41:20] please. [41:20] >> Yes, please do. It's it's excellent. Uh [41:22] there was actually another part of that [41:24] question um talking about um uh using it [41:27] in VS Code. Can it be used in VS Code [41:30] Copilot? Uh so yeah, this kind of gets [41:33] difficult to um explain. Um but here's [41:36] here's something that I'm going to do. [41:38] So, I'm going to open up this um I'm [41:40] going to do it in another window so I [41:42] can open [41:45] um [41:46] a different repository. [41:50] And now we're going to All right. So, I [41:55] have opened VS Code here and I've opened [41:58] it in a personal project of mine. Um [42:02] it's a big old Java code base. Um and [42:08] We're going to go into the terminal. [42:17] This is where we learn if my VS Code [42:20] config is broken from me testing things. [42:22] Okay, no, we're good. Uh, but if I do [42:24] copilot, [42:30] it's going to load my environment. Um, [42:32] and you'll notice that here it said uh [42:33] Visual Studios Code Insiders connected. [42:36] Um and so now the we have integration [42:40] with um with VS code and so if I do [42:44] things like um uh I don't have a very [42:49] let's I don't know add a command [42:53] to I don't know this is this is this is [42:56] a big Minecraft uh uh server plug-in uh [43:00] codebase is what this is. Yeah. So I'm [43:01] trying to think of something simple. I [43:03] don't know. add a command to give the [43:06] user 64 diamonds. Um, and so when it uh [43:13] it can grab context from the VS Code [43:14] session, so it can look at, you know, [43:16] the files that you have opened uh and [43:17] all of that. Um, but then also when it [43:19] shows diffs, uh, it can show them in the [43:21] VS Code diff viewer um, as well. And so [43:25] it it kind of is you can use the copilot [43:28] CLI um if that's the harness that you [43:30] prefer or if maybe um you know you just [43:34] uh um you know want this kind of [43:36] continuity between you know the sessions [43:38] that you're working on in VS Code and [43:39] not in VS Code um you can um [43:44] you can spin up uh the the the Copilot [43:47] CLI in VS Code um to then talk to uh you [43:51] know this harness that you already know [43:52] and love. [43:54] I love that. And so now I have all the [43:56] context of my project plus what I have [43:59] open. [44:00] >> Exactly. Yeah. Yeah. Yeah. Um I'll be [44:03] totally honest. I haven't used this that [44:05] much because again um I have just uh [44:07] like you kind of just completely taken [44:10] the editor out of my life and I've gone [44:12] full uh I've gone full terminal. Uh and [44:14] that's the way that I like it. Um but uh [44:17] yeah, so it's it's spinning up a ton of [44:18] explore agents here because um uh uh I [44:22] think I have an old build of this plugin [44:25] where I don't have um I don't have any [44:28] of my my my agent uh instructions set up [44:32] >> and so that's why so it's going through [44:34] um but like you know it is remarkable [44:35] like you know the fact that you can spin [44:37] up the explore agent you know multiple [44:39] parallel explore agents to go and figure [44:41] out how this process works um and it is [44:43] just going to at the end of this um just [44:47] um you know kind of figure out how this [44:49] codebase is laid out, find where the [44:51] commands are, find how they're [44:52] structured um and then and then and then [44:55] write a command for me without ever [44:56] having you know like seen this before [44:58] and that's that is a very uh very [45:00] magical thing [45:02] >> and there is it's going you say more [45:04] about this running task in parallel um [45:06] because I think that's something that we [45:08] don't talk about enough right and then [45:10] you have to show us fleet too because I [45:13] think that's one of the most powerful [45:14] things that like this is why this is the [45:16] harness you need. [45:18] >> It's true. It's true. Yeah. So, running [45:19] things in parallel is a is a superpower. [45:22] Um, you know, it used to be that you had [45:24] your agent do one thing uh and you said, [45:26] "Here, go off and build this [45:28] functionality for me. Go off and change [45:29] this." Um, and uh, you know, it went off [45:32] and did it and you couldn't have it. You [45:34] know, who knows how long that was going [45:36] to take, right? It might take a lot of [45:38] uh, a lot of, you know, retrying certain [45:40] things. it might take um uh might be [45:43] just like a lot of steps and it just [45:44] executes all those steps one [45:46] sequentially after the other. Um but uh [45:49] nowadays what we really want is for um [45:52] the agent to create a to-do list of all [45:55] of the tasks that it needs to do figure [45:57] out okay these things can be done in [45:59] parallel and then do that in parallel. [46:02] Uh okay finally so now we got so many [46:05] things happening at once. Uh so now uh [46:08] all those um parallel explore tasks [46:11] finished. It figured out how it wanted [46:12] to add a new command and now it's [46:14] showing me the diff in VS Code's diff [46:17] viewer. Um so I can uh of course it's [46:20] less exciting when it's creating a new [46:21] file. But you could imagine that if this [46:23] was a diff of an existing file. Uh you [46:25] know it would be uh better to see that [46:27] in here than um uh than than in the uh [46:31] in the IDE. And so then instead of [46:33] saying yes in the CLI, I can say yes up [46:37] here. Um and now it can go off and uh [46:40] and do that. So that's that's pretty [46:42] cool. Um that is cool. [46:44] >> Uh and here you go once again. Now it [46:46] needs to register the command and it [46:47] shows it there. And I can say yes up [46:48] here again. [46:49] >> Um but fleet. [46:51] >> So with Fleet when I and I don't think I [46:55] even have Java and Gradel installed on [46:57] here, so I don't think that build's [46:58] going to work. So for let's just cancel [46:59] that. Um, if I have a sufficiently large [47:03] task that I want the um the model to go [47:06] off and figure out, it's going to create [47:07] a to-do list and it's going to figure [47:09] out um here are all the constituent [47:10] parts of completing this task for you. [47:13] >> And then um it is going to, you know, [47:16] again figure out the dependency order of [47:18] these tasks, figure out which ones can [47:19] be done independently from one another [47:22] and then dispatch a several parallel [47:25] agents each with their own context [47:27] windows, right? in each with their own [47:29] um tasks from the to-do list to do um [47:32] and and go off and and complete each of [47:35] those um those bits of the tasks. Um and [47:37] if you combine this with autopilot mode [47:40] um which is essentially yolo mode and [47:43] then also just constantly looking at the [47:45] to-do list saying is there anything left [47:46] to do? Is there anything left to do? Is [47:48] there anything left to do? Until it's [47:49] completed. Um you really just have um [47:52] you know a a true fleet of agents that [47:54] can go and um and complete a task. um uh [47:58] completely completely autonomously [48:01] >> and you don't have to assign specific [48:05] like you're not spinning up these agents [48:06] just using the slashfleet command takes [48:09] care of that for you. [48:10] >> Exactly. Exactly. And so my loop is make [48:14] a plan, get the plan, um ask it to fire [48:18] up parallel sub agents of Opus, uh [48:21] Codeex and Gemini to review said plan. [48:24] Then once I like the plan, I go into [48:26] autopilot mode. I say slashfleet [48:28] implement the plan. And then it just [48:30] goes off and and does that. So let's [48:31] let's let's do that right now. Let's go [48:33] into plan mode. Um and let's say we want [48:36] to [48:38] um [48:40] I don't know, let us uh add a [48:45] tag [48:46] miniame [48:48] uh party game. So, the project has this [48:51] notion of party games that people can do [48:53] if they're in a party on the server. Um, [48:56] I'm going to have it make a tag miniame [48:58] uh for party games. Um, so it's going to [49:01] go and create a plan. And so, it's going [49:02] to do that by exploring how the system [49:04] works using the explore agent that we [49:06] had used before. Um, so it's just [49:08] looking, you know, it has no clue how [49:09] any of this works. It doesn't have very [49:11] good extend uh uh custom instructions [49:13] for this codebase. Um, but you know, [49:15] it's it's okay, it's found a pointer, so [49:17] it's going to go and do that. Um, and [49:18] then it's going to create its plan. And [49:20] then once the plan's created, I'm not [49:21] gonna iterate on the plan. We're just [49:23] gonna have it go so we can see what [49:24] fleet mode looks like. [49:26] >> Let's yellow it. I love that. Can you [49:28] can you say a little bit about how like [49:29] at this point even during the [49:31] exploration like I guess the value of [49:33] having co-pilot instructions in there? [49:35] Um [49:37] can you guide how that exploration is [49:39] done? Is it nate to the harness itself [49:44] or can we actually kind of guide that [49:46] because we all have sort of different [49:47] ways that we start kind of like the [49:49] thinking process? [49:51] >> Yeah. Yeah. I mean for me what um I I [49:54] think [49:56] >> if even if you don't have any [49:57] instructions [49:58] >> the agent is remarkably good at scanning [50:00] the code base for you know your design [50:03] patterns and figuring out kind of you [50:05] know a pretty decent way of uh of [50:07] working at it. Um but what the custom [50:09] instructions are going to do is it's [50:11] going to give the agent [50:14] um a very baseline view of you know hey [50:16] here's the programming language that [50:17] we're using. here's like the the the [50:19] main architecture of this and the [50:22] constituent parts of it. So that way [50:24] when the explore agent starts, the first [50:26] thing it's doing isn't looking for yml [50:28] files, which was the first thing it was [50:30] doing here, right? It's like, hey, let [50:31] me look for the word party. Mml [50:33] anywhere. It's like, well, it's not [50:34] going to find that. I know that because [50:36] I've worked on this codebase for 10 [50:37] years, but like, you know, like it [50:39] doesn't know that because it was never [50:40] told that. and and and that's what the [50:42] custom instructions are going to um to [50:45] help it get through a little bit more um [50:48] a little bit more easily. Um [50:52] definitely not necessary. You definitely [50:53] don't need it. The agent's still going [50:55] to do a very good job at figuring out [50:56] how to use your codebase without it. Um [50:58] but it shortcuts the process. It'll make [51:00] uh you know the the the the part that's [51:02] happening right now um you know happen a [51:04] little bit uh a little bit quicker. [51:06] >> Just a little bit quicker. Yeah. And [51:07] then again, if you start thinking about [51:09] it, I mean, that's one thing because [51:10] remember, we're doing this very much [51:12] like you will develop software. Right [51:14] now, you're in the exploratory phase. [51:16] You're figuring out, you have your idea, [51:18] you want to add this party game, and you [51:20] have this codebase, and now you just [51:22] kind of figure out how that's going to [51:23] fit in there. Uh, and then, you know, [51:25] we'll move on from that. But if you [51:28] think about skills, this is one thing I [51:30] feel like is a question that comes up a [51:31] lot, like skills versus agent. like when [51:33] would you use a skill versus just [51:35] creating that custom agent for a job [51:37] that you do all the time? [51:39] >> Yeah. Yeah. Yeah. To to me an agent is a [51:44] type of [51:46] a type of person with a certain kind of [51:48] like a remmit, right? So like you know [51:51] you think of like um um like our [51:53] built-in agents. We have an explorer [51:55] built-in agent. We have a review [51:56] built-in agent. these are these are like [51:59] you know like like it's a specialized [52:01] person who knows how to explore things [52:03] really well and here's the list of tools [52:04] that I have that allow me to explore [52:06] things really well and a skill like are [52:08] those tools that you give those people [52:10] >> beautiful [52:11] >> right and so you create kind of like a [52:12] persona of here is the type of person [52:15] and what they do to um uh to help me [52:18] work on my project and then skills are [52:20] the actual tools that you give them um [52:23] and the actual skill set that they have [52:24] in order to do a task [52:27] Perfect. That's beautifully explained. [52:29] Okay. So, he's going he's going while [52:32] it's cooking. I want to I want to go [52:34] through the questions that are coming up [52:36] here. I love this question because I'm [52:38] like use what makes you happy. But [52:40] honestly, like out of all people in the [52:42] world because I'm a huge insiders chill. [52:44] Like I love VS Code Insiders. It's been [52:46] my my working editor forever. But when I [52:48] tell you I haven't opened that bad boy [52:50] for weeks, there is a reason. Uh why do [52:54] you prefer I mean why do you prefer like [52:58] >> it's a personal choice but also [53:00] >> can we say about how I guess it's [53:02] different? [53:03] >> Yeah. Yeah. I mean it is it is certainly [53:05] a personal choice and I would say that [53:06] I've always been the type of person who [53:08] prefers to be in the terminal um if I [53:10] can um and and and that's just a reason [53:13] to do it. And I think that's again part [53:15] of the reason why we have both of these [53:17] products is because it does allow you to [53:19] kind of uh to to to choose. Um, but I [53:22] think the CLI, [53:25] part of it is, uh, we've gotten to some [53:27] cool UI here, so we'll get to that in a [53:30] second, but [53:30] >> yeah, I like that. [53:31] >> Uh, yeah. Um, but the CLI, [53:36] >> it's a method of working with a computer [53:39] that is um, obviously easier to stick [53:41] into um, automation. And so that's [53:44] obviously one big reason why we have [53:45] this. And maybe that isn't relevant to [53:48] to you and and and your daily life, but [53:51] um you know like that's one use case for [53:54] having it. But another one is well I [54:00] mean [54:02] we've seen like with MCP servers and [54:05] with agent skills that having structured [54:08] ways of agents to be able to operate [54:10] software is really important. And the [54:13] copilot CLI is now a way that agents can [54:16] structured interact with co-pilot, [54:19] right? And so like we have the copilot [54:20] server mode so you can have your [54:22] applications actually like drive a [54:24] session with copilot. You you not going [54:26] to be able to do that just with the side [54:28] panel in the IDE because you can't [54:29] programmatically access it. It's all [54:31] like you know enabling programmatic [54:33] access. Um and so like yes are the [54:37] features you know are there features the [54:38] CLI has that VS Code doesn't have? Are [54:41] there features that VS Code has that the [54:42] CLI doesn't have yet? Um, yes, like that [54:45] that is that is true, but it's a [54:46] different method of interacting with a [54:48] computer um that you know kind of allows [54:52] you to pick and choose what the [54:54] environment is going to um going to have [54:58] um and makes a little bit easier to uh [55:02] um a little bit easier to interact with [55:04] it in like you know headless [55:05] environments. Um, you know, if I if I [55:07] SSH into a computer that I don't access [55:10] very often and I don't have a monitor [55:11] attached, I can install a copilot CLI [55:14] and now I have an assistant in that [55:15] terminal ready to go and ready to help [55:17] me uh, you know, administer the system [55:19] that I don't touch very frequently. Like [55:22] what am I going to do? Install code [55:23] server and then uh, you know code server [55:26] in the browser. [55:26] >> Try and run a uh a a remote like a [55:30] pretend VS code online so that you can [55:32] do something that you can do from the [55:33] terminal with copilot. No, that makes [55:35] total sense. That makes total sense. [55:37] That's what [55:38] >> but but maybe it matters to uh to this [55:40] commenter. I forgot their name, [55:40] unfortunately. But maybe it matters to [55:42] them that they do they they would rather [55:44] do that and they would rather, you know, [55:45] have the the GUI there and do it that [55:47] way. And that's okay. And I'm not going [55:49] to sit here and tell you that you must [55:50] use the CLI and only use the CLI and [55:52] that's the only good way to do it [55:53] because that's not true. [55:54] >> Yeah, it it really is about giving [55:56] everybody uh choices, [55:57] >> having the choice honestly. And there [55:59] are things that it depends of what like [56:01] I was just in the process of setting up [56:04] a new computer and I neglected doing it [56:07] the right way and like saving all my [56:10] config files and like doing and so [56:12] before I moved from one place to the [56:13] other I use copilot I kicked it off on [56:16] the CLI but then I actually wanted to [56:17] look through the files and a little bit [56:19] more in depth way. So I'm like okay [56:20] let's open this in and in and insiders. [56:22] So there is I mean I think the advantage [56:25] here is that you can choose when and [56:27] how. Um, so it's it's it's your it's [56:30] your choice. Can you tell us what's [56:32] going on? Because look at this. [56:35] >> This is great. There's been so many um [56:37] UI advancements in the C-pilot CLI [56:40] recently and you know it's sometimes [56:41] tough for me to keep up with the channel [56:43] where the designers are just [56:44] experimenting with stuff and coming up [56:46] with cool ideas. Um, and so, uh, when [56:49] you're creating a plan especially, you [56:51] know, it's really [56:52] it's really important that, um, that [56:54] that you are answering clarifying [56:56] questions that the model gives you. And [56:57] we noticed this a lot. And, you know, we [57:00] were just like, you know, the model [57:02] would ask 10 questions. You would copy [57:04] each question and say, number one, yes, [57:06] this, number two, I think you should go [57:08] with this option. Um, and that was [57:10] really cludgy. That was really slow. Um, [57:13] and probably not a great use of tokens. [57:15] And so what we did was is we created [57:17] this ask user question experience and [57:18] we've made lots of um different uh [57:20] iterations to it already. Um but here [57:23] we're building this tag miniame [57:25] and uh you know copel wants me to know [57:28] to to tell it what uh style of tag is it [57:32] classic tag where one player is it and [57:34] you tag someone to pass it on? Is it [57:35] infection tag? Is it freeze tag? [57:37] >> I love this like elicitation style of [57:40] like working is beautiful. [57:42] >> Yeah. Yeah. And so we're and this is [57:43] this this and this new elicitation style [57:45] we're also um expanding to MCP server [57:47] elicitation as well because that's that [57:49] that's that's not in yet but um we've [57:51] gotten a lot of feedback to uh to enable [57:53] that and this [57:54] >> this framework is going to allow us to [57:56] do that as well. Uh so we're going to go [57:57] with classic tag. Okay. Um how long [58:00] should a round last? Uh and it's filled [58:02] in 120 seconds um to be like a you know [58:06] a starting point. Um that's fine. Sure. [58:09] Um how many players minimum? I guess two [58:13] two player tags is kind of boring. Let's [58:15] Let's start it with three. I think I [58:16] think you need at least three people for [58:18] tag to be uh truly uh interesting. Um [58:21] maximum players. I don't know. Let's do [58:23] 20. Um what warp or server location [58:27] should this game use? [58:30] >> Um that's cool. So like where should the [58:31] game start? And that's fine. Let's let's [58:33] let's do that. Let's start it um there. [58:36] Um how do players tag each other? Um, [58:39] let's punch or hit them. [58:41] >> Okay. [58:42] >> Any other details or notes? I don't [58:43] know. No, I'm good. And I could do [58:45] control S. And so now I've just answered [58:48] all those questions. So I do control O [58:50] here to look at the um uh Oh, never [58:53] mind. It's not actually going to expand [58:55] that. Um, but you can see the beginning [58:57] of it here. Like you know, it asked [58:59] questions. So variant equals classic, [59:01] duration equals 120. And so it has all [59:03] the answers to that now. And it's going [59:04] to use that to create the plan. Um, and [59:07] then we're I know we're we're we're [59:09] almost at our scheduled end time, but I [59:11] do want to kick off [59:12] >> cooking. No, we're gonna watch it cook [59:14] if you're okay with it. [59:15] >> Oh, I'm more than okay with it. [59:16] >> Okay. And then see you here. What this [59:19] just facilitated is you're saving on [59:21] token. This would have been a [59:22] conversation that would have been back [59:23] and forth, but the CLI was intelligent [59:26] enough to like say these are decisions [59:28] that we need to make before I go bake [59:31] this thing. and presenting them in a way [59:33] that is like just very practical. [59:35] Beautifully designed. By the way, the [59:37] team does the sign in this is [59:39] >> it's excellent. It's excellent. Um ever [59:41] ever since we came out with uh with that [59:43] with that with the uh co-pilot banner, [59:46] um you know, we've been we've been [59:47] cooking with design. [59:48] >> Beautiful. I'm going to play that just [59:50] cuz I love it so much. [59:52] >> Same same. I'm a big fan. I'm a big fan [59:54] of the Oh, beautiful. Look at that. It's [59:56] the little things, friends. Okay. It's [59:58] the little things. [59:59] >> It's true. All right, let me take a [60:01] quick look through the through the chat [60:03] to make sure that how is Copilo so [60:06] cheap? [60:07] You mean free? Like it's free for you. [60:10] You can go use the CLI right now. [60:12] >> Um, [60:13] >> please do. Please do. [60:14] >> Please do it. Let's see. And [60:18] uh I still don't understand tokens. We [60:20] can talk and honestly like I think [60:21] that's a larger conversation. Ryan, [60:22] maybe you can come and play with this. [60:23] I've been talking to my team about doing [60:25] a stream that's about that like helping [60:27] people guide their options with when [60:29] they're modeling because we live we sit [60:31] in our ivory tower of unlimited tokens. [60:33] >> It's true. It's true. [60:34] >> But for our own personal projects and [60:36] other things like it's good to think [60:38] about better ways to efficiently deploy [60:40] those agents and do all the things. [60:41] >> Yeah. Yeah. I mean and the beauty of [60:43] co-pilot is you know that um you know we [60:46] do charge by premium requests and not by [60:48] token right. So there's a lot of things [60:49] that you know yeah you know who cares [60:51] how many tokens you're using. Uh but [60:53] like you still should care how many [60:54] tokens you're using. Um [60:56] >> for for one reason uh you know like the [60:59] model does have a finite context window, [61:01] right? And as you as you get to the end [61:03] of that context window um we've done a [61:06] lot to um again have auto compaction [61:08] happen. And so by default, you know, [61:10] when you get close to to to exhausting [61:12] your context that we will automatically [61:14] compact it uh so the model only keeps [61:16] the most relevant things in its context [61:18] window. So you can continue that [61:19] session. Um and then also with copilot [61:21] memory and with um uh you know the SQL [61:24] database that um that we've given the [61:25] agent so it can have its to-do list and [61:27] everything. Um [61:30] uh it um [61:32] it can have the session go on forever. [61:34] Um but still, you know, if it can do [61:36] that fewer times um that's better. Uh [61:39] and so, you know, you still probably do [61:41] want to um to conserve your context [61:42] window and conserve your tokens if you [61:44] can. Um but here we go. Okay, our plan's [61:46] ready. Um [61:47] >> Whoa. Okay. And so I can uh do control E [61:51] to view the full plan. So here's the [61:53] full plan that we have here. Um [61:57] sounds good to me. Um [62:01] now I could just say accept plan and [62:03] build on autopilot and then it will just [62:04] automatically move me into autopilot and [62:06] start building this. Um, I'm going to [62:08] prompt myself because [62:11] um, we want to go into [62:14] um, autopilot mode now that [62:16] >> So, is autopilot the same as YOLO? And [62:20] there was a question about YOLO [62:21] actually. Does it mean stop asking me or [62:23] does it mean just [62:25] >> go to town? [62:26] >> Yeah. So, YOLO is the combination of [62:29] allow all tools and allow all paths. So, [62:32] by default, when Copilot wants to um [62:34] access a path that isn't in the current [62:36] working directory, you get prompted, [62:38] hey, can co-pilot go look at this path. [62:40] That's very good security practice, [62:42] right? You because you might not want [62:43] Copilot to go and be messing with your [62:45] system files or anything like that. Um, [62:47] and then allow all tools, I think, is [62:48] pretty self-explanatory. You know, [62:50] you've seen Copilot ask for um [62:52] confirmation to run tools all of the [62:53] time. Um, so YOLO is the combination of [62:56] allow all tools and allow all paths. [63:00] Autopilot is the combination of YOLO and [63:04] don't ask me follow-up questions. Um, [63:07] yeah. And so that combined with the [63:10] model's to-do list allows it to just add [63:12] things to the to-do list and then just [63:14] constantly just keep looping and looping [63:16] and looping until everything has been [63:17] completed. And so it gives the model [63:19] this tool to say like, "All right, just [63:20] keep going in a loop until" and then [63:22] when you think you're done, call this [63:24] I'm done tool and then we can go back [63:25] out um to have the user do stuff. Um, so [63:30] maybe a little confusing there, but [63:32] that's kind of how YOLO and Autopilot. [63:33] There's actually a really good article [63:35] on the docs about um, uh, let's see, [63:39] Copilot CLI. [63:41] >> Let's share that link, please. If you [63:42] pass it to me over here in the private [63:44] chat, I'll I'll I'll share the right one [63:45] because that will be [63:46] >> I will um, [63:48] >> that will be very very useful. [63:52] >> Yeah. So anyway, so so I'm going to do [63:54] slashfleet implement the plan here. Um, [63:56] and what that's going to do is start a [63:57] fleet of parallel agents that are going [63:59] to take that to-do list and figure out [64:01] what can be run independently of one [64:03] another and then run all of those um uh [64:06] uh in parallel. Uh, and then I'm in I'm [64:08] in autopilot mode. So, it should just [64:09] keep going until the plan is complete. [64:11] And so then I can just uh say that uh so [64:14] of course I'm in autopilot mode. So we [64:16] have to make sure that um we give it all [64:18] permissions. And I will say yes. [64:20] >> Okay. [64:21] >> And it's implementing the plan. And as [64:23] we do that, I'm going to get this um [64:26] uh this link all about [64:30] autopilot mode. There's that. So share [64:32] that with the folks if they want to go [64:33] learn more about it. [64:34] >> Thank you. For sure. Um and that was a [64:36] really good explanation on on the [64:39] differences between YOLO and autopilot. [64:42] Now with the fleet command, you just did [64:44] slashfleet and what was your prompt or [64:46] there was no you just said go implement [64:48] this plan. [64:49] >> I said implement the plan. Yeah, [64:50] >> but you could have been more granular, [64:52] right? Like you could have been more [64:54] specific. [64:55] >> Yeah, I could have said just implement [64:56] phase one of the plan and then come back [64:58] and we can talk about, you know, kind of [64:59] how that went um ahead. I could say [65:01] implement the plan with changes and tell [65:03] it what I thought the changes I wanted [65:05] to make. Like, you know, it it just is [65:07] it is natural language. I could have [65:08] told it to uh to do whatever um whatever [65:11] I want it to do. [65:13] >> Perfect. Look at how it's spawning its [65:15] own little agents to go to work. [65:17] >> Yeah. Uh, and so, you know, it's it's [65:18] creating the events that it needs. Uh, [65:21] it's creating the what whatever scoring [65:24] looks like in tag. What does scoring [65:25] look like in tag? I probably should have [65:27] read the plan, then I would have known [65:28] uh what that was going to be. But, um, [65:30] so yeah, you know, and then so then it's [65:32] spinning up these agents to go through [65:33] and do that. So, it has two independent [65:34] to-dos, I guess. Um, one was the event [65:36] and one was the score. And that's what [65:38] it decided to do. Um, so all right, now [65:40] that it's done that, it can um do the [65:42] next two uh agents. So, one is actually [65:45] creating the notion of the tag game and [65:47] the other one is um listening for when [65:49] players punch each other because that's [65:51] how we're going to know that somebody [65:52] has been tagged it. Um and so, you know, [65:55] and it's just going off and and doing [65:57] its thing. [65:59] >> Amazing. [65:59] >> And we could go we could go watch it if [66:01] we wanted to and all of the uh and all [66:03] of the changes here. Um but yeah, it is [66:07] uh it's pretty it's pretty powerful. Um [66:10] >> very cool. You can go from literally [66:12] just an idea to having working software [66:14] just uh with a few uh a few back and [66:16] forths with the model and that's cool. [66:18] >> Okay. Can we talk about the resume as [66:20] well? Because that's another thing [66:21] especially for us today like we were [66:23] cooking but we got to go feed the dog [66:26] and go for a walk. Um so if I'm sort of [66:29] not finished with my session then [66:32] >> yeah. So, one thing that uh you can do [66:35] is if I'm just in a session and I just [66:36] say hi. Um [66:39] whatever [66:41] we'll let that do that real quick. Um [66:44] >> and after it comes back, if I close the [66:47] session, uh you'll get this in chat. So, [66:49] this will give me a big long um UU ID. [66:52] Uh and if I run this command, then I can [66:56] resume that specific session. And so, I [66:58] can go look at back in my in my terminal [67:00] history and find the specific session I [67:01] was doing. That's not a great experience [67:03] though. Um, and so what you probably [67:05] want to do is just do copilot-res. [67:08] And that is going to spin up um a very [67:12] cool um [67:15] menu. [67:15] >> Look at that. [67:16] >> Yeah. [67:17] >> With a summary of what the work was. [67:19] That's brilliant. [67:19] >> Exactly. Yeah. And so we can go and we [67:21] can look at um so these are the ones [67:23] these these ones at the top were the [67:25] ones that we had in this um git uh [67:28] repository. Um but then here are just a [67:30] bunch of other sessions too. And so we [67:32] can go and find the one that we want to [67:33] uh that we want to do. And you'll notice [67:35] up here um sessions local and remote. [67:38] The remote sessions [67:40] I can resume a session that I had [67:43] running in the cloud with Copilot coding [67:45] agent in my terminal. And that's pretty [67:49] magical. So we can both delegate a [67:52] session that I'm working on in my [67:53] terminal up to copilot coding agent. So [67:55] that way I can just be working on it in [67:56] the cloud. But I can also then take the [67:58] session from copilot coding agent and [68:00] resume it here in my terminal locally. [68:02] >> So what that magic means is that you [68:05] could assign a task to the coding agent [68:07] when you leave your house to go walk the [68:09] dog [68:10] >> and then when you come back and you're [68:11] ready to dig into the work, you can go [68:13] ahead and resume that session from your [68:15] terminal or from if you're using the ID [68:17] from your ID. [68:18] >> Yeah. Yeah. Or you could be crazy like [68:20] me and actually just have a terminal app [68:21] on your phone and just literally run it [68:23] on your phone while you're walking the [68:24] dog. You could do that too. Say more. [68:27] Say more. What are you What are you [68:28] running? How are you? Tell me later. [68:31] >> Yeah. Yeah. Yeah. Yeah. I'm I'm on [68:32] Android. There's some terminal apps on [68:34] Android you can use. They're very cool. [68:35] Um yeah, [68:36] >> that is amazing. Question about the [68:38] models and I've done this experiment [68:40] actually. Uh does it use the same base [68:43] model and base agent? So how can you use [68:46] custom agents when you're using fleets [68:48] or toy set of task in parallel? And [68:50] thank you so much for the question. [68:51] >> Yeah, thank you for the question. Uh [68:52] yes, you can. And so I'm going to scroll [68:54] up a little bit here. And you'll notice [68:55] that um when Fleet was dispatching its [68:57] agents, it's it's creating general [68:59] purpose agents. That's our that's just [69:02] our built-in agent. Um if I had uh if I [69:04] had a bunch of custom agents in here [69:06] that I had with, you know, say, oh, I [69:07] want to make sure it's using this model. [69:08] I want to make sure that it's um uh you [69:11] know, only has access to these tools or [69:13] whatever. Like um yes, you you you could [69:16] prompt it in your slashfleet. You say [69:17] slashfleet [69:19] use these agents to uh to complete these [69:22] tasks. and instead of the general [69:23] purpose agent, it will uh use those [69:24] agents. Um, or I bet you could even if [69:27] if in your custom instructions you said [69:29] this is the agent that you use for these [69:31] kinds of tasks, I imagine that it will [69:33] intelligently select the right agent for [69:35] the job there as well. Um uh but even if [69:39] you don't have any custom agents defined [69:41] um it will also so here if I um let us [69:47] go back to here if we go to uh let's [69:51] resume [69:54] um [70:00] I can um tell it to I don't even need to [70:03] be in fleet I could say spin up parallel [70:06] sub agents [70:08] >> using codeex 5.3, [70:12] opus 4.6, [70:14] and Gemini 3. Is it 31? Is it three? I [70:17] think it's three. Um, I think three is [70:19] the one in three. Yeah. [70:21] >> Um, uh, to review our changes. Um, and [70:27] it has the ability to just kick off a um [70:31] a general purpose agent using these um [70:34] individual models as well. Um, so if [70:37] it's just if you don't want to have a [70:39] whole custom agent defined and you just [70:40] want to use a different model um for [70:42] your for your sub agents, uh, it can [70:44] handle that on its own as well. [70:46] >> And this is the this is the sauce right [70:48] here, friends. This is how you get that [70:51] robust opinion baked into the work that [70:54] you're doing because you're you're [70:55] you're using multiple agents for that. [70:58] >> Ben, the remote feature would be resume. [71:01] I mean, because you can pick up the [71:03] session or I don't know if if you can [71:04] say more about that, right? [71:05] >> Yeah. Yeah. I mean, um, that's the [71:08] beginning, I think. Right. Um, I think [71:11] there's more in that space that we'd [71:12] like to do, especially after seeing what [71:14] some of our friends, uh, who who make [71:16] some other agent harnesses have done. [71:18] think I think there's some cooler stuff [71:19] we can do in that space and uh we're [71:22] we're we're looking into it. We're [71:23] looking into it. [71:24] >> Okay. Yeah. But again, if you have tried [71:27] the copilot CLI in September and you [71:29] have not tried it since, please upgrade [71:32] to the latest version. You can do [71:34] install it using brew. If you're on [71:36] Windows, I think there's a wing um [71:38] installation. It it works in Mac OS, [71:42] Windows, Linux. It's free. [71:46] >> It is. Again, it cost you [71:48] >> zero dollars. There is the installation [71:51] guide and there is an entire and growing [71:54] ecosystem of plugins and skills and the [71:56] also co-pilot repository. I will share [71:58] the link to that as well. Uh because [72:01] definitely you have to try it out. Ben, [72:04] yes, Ben, I'm here for you in this [72:06] question because this is me. I want to [72:07] be able to go outside. [72:09] >> Oh, yeah. Yeah, we we we're working on [72:12] it, Ben. We are working on it. Stay [72:13] tuned. [72:14] >> Stay tuned. I think like you said um [72:17] that was the beginning. So and the way [72:19] that how fast this team is cooking I [72:22] don't think it'll be too long. So [72:24] amazing amazing uh you can install it [72:27] today. You can start trying it out. [72:29] >> Can you tell us a little bit about [72:32] >> what's okay maybe maybe one more [72:34] question about context because this is [72:35] and this is such an excellent question. [72:37] Thank you so much for asking that you [72:39] betray. If you have an app that involves [72:41] multiple repositories, each one its own [72:43] project, have you got a specific [72:46] strategy that's worked for you, like a [72:47] best way to develop it on a efficient [72:50] manner? Uh [72:52] >> yeah. Yeah. I mean, what I do, and I'm [72:54] not going to pretend like this is [72:55] state-of-the-art, and this is the best [72:56] possible thing that you can do, but what [72:58] I do, and it works for me, is I open [73:01] Copilot in a directory that just [73:03] contains all of the um the sub projects [73:06] as as as subdirectories that I care [73:08] about. Um, and then I just I just I just [73:11] tell it here like here's each of these [73:13] different um repositories. Here's their [73:16] concerns and how they fit into the [73:17] bigger picture. Um, and then the the [73:19] agent is just smart enough to go in [73:21] there and um and and know um which repos [73:25] to to to go into for for different [73:28] tasks. [73:29] >> Oh, absolutely. Absolutely. There's [73:31] definitely some workflows that I've seen [73:32] out there um that uh get a little bit [73:35] more, you know, crazy with how they [73:37] define custom agents for working um in [73:39] in each of the repos, but um you'd be [73:42] surprised at [73:44] how little of that special tooling you [73:46] really need um to to get good results. [73:49] Um um and you know, if you if you want [73:52] to move to the monor repo life, uh you [73:54] you most certainly can [73:55] >> come to the dark side. But honestly [73:57] though, I have been so impressed because [73:58] I've done a lot of this kind of cross [74:00] tasking um and it's handled it [74:02] beautifully where it's like I'm doing [74:04] something and I'm like oh by the way in [74:05] this other repo because guess what [74:07] friends, the GitHub and CP server is [74:08] baked in it. you can see all my all my [74:11] repos. You know, you know what I'm [74:12] talking about when I tell you go look at [74:14] that other project and tell me how we [74:16] did especially experimenting with the [74:18] copilot SDK which we didn't even get to. [74:21] Um because you know just doing some kind [74:24] of advancerish workloads for different [74:26] experiments I was running and I'll be [74:27] like well it work on there go see how [74:29] you did it there and then come back and [74:33] it just it handled that. I'm not running [74:35] work trees or no, I'm just having one [74:38] conversation, but the agent's smart [74:40] enough to figure out what projects and [74:42] and where things need to go. So, that's [74:44] worked out beautifully uh for me. But I [74:46] love I love the disclaimer. You just [74:48] kind of kind of experiment with it. Um [74:50] again, it cost you zero dollars, so [74:52] experiment with it and you'll figure out [74:53] the best way to to work with it for you. [74:56] Uh let me take a look and see if there [74:58] are any other questions. I I do want to [75:00] see this question about Hello, Mr. Cojo, [75:04] friend of GitHub, friend of the show. [75:05] Um, the coolest feature that you have [75:08] come across so far, copilot CLI, or [75:11] what's a thing that you've done in the [75:12] CLI that you're like, you already showed [75:14] me today Chronicle and Chronicle Tip [75:16] that I have not played with. Uh, fleet [75:18] and delegate are they're my friends. I [75:21] mean, we've been we've been there hand [75:23] in hand and that's how we work. being [75:25] able to uh resume the sessions and [75:27] having that slug where it's actually [75:29] giving you a summary of what it was [75:31] worth so you know exactly what session [75:33] to pick. Uh amazing. What else have you [75:36] seen out there or that you've been using [75:38] yourself that you're like well that's [75:40] cool. [75:40] >> Uh yeah. Um well okay my my true answer [75:43] to this question is Chronicle. I think [75:45] right now part of that is because of how [75:46] new and how and how cool it is. Um, but [75:50] oh boy, I think the the coolest use case [75:54] that I've seen um is [75:58] people creating their kind of, you know, [76:00] open claw like personal assistants and [76:02] going crazy with kind of, you know, [76:04] organizing their entire professional [76:05] lives around this tool with a million [76:08] different um uh with a million [76:10] different, you know, custom plugins and [76:12] skills and everything. Uh, and I know [76:13] Patrick showed you his thing where he [76:15] has Copilot call him when it needs help [76:17] with things. [76:18] That's really cool. Uh then that that [76:20] that's magical every time uh that I see [76:23] it. Um [76:25] >> but it, you know, it just really goes to [76:27] show that there is no limit on what you [76:30] can do with this tool. Um you know, you [76:33] have the building blocks to truly kind [76:35] of do whatever it is that that that you [76:38] want. Um but for me, for me honestly, [76:40] like it is it is chronicle. is the fact [76:42] that I have a way to uh learn how to use [76:45] the product better and collaborate with [76:47] the agent better and use it more [76:49] effectively um from within the product [76:51] and you know not everyone needs to be [76:54] like me and in the the Slack channels [76:57] with the engineers as they're coming up [76:58] with stuff and they have their little [77:00] tips and tricks like I can have the the [77:03] master of using the co-pilot CLI um in [77:06] my terminal with me at all times and ask [77:08] it questions about how to get better. [77:10] >> Absolutely. It's magical. It really is. [77:12] >> And so much of that is like your own [77:14] personal experience and meeting [77:15] developers where they are. Like we want [77:17] you to use the tool the way that you [77:18] want to. And in order to do that, like [77:20] you have that flexibility now. Okay. You [77:23] have to tell us what's coming up. I know [77:26] you have a you have a list of requests. [77:29] >> Uhhuh. [77:29] >> Oh, hello. We can also provide feedback [77:32] right from the CLI. [77:33] >> That's true. [77:34] >> Yes. Yes. Yes. Oh, is it finished? [77:36] >> It's finished. Yeah. So, our our tag our [77:38] tag game completed. And uh not only [77:41] that, but um oh well, the code review [77:44] got stuck. I didn't run my my review on [77:46] autopilot, so it it stopped asking me to [77:48] run Gradle. And it's not even going to [77:50] work because I don't have I don't have [77:51] all that set up. Uh but yes, the tag [77:53] game did complete. [77:55] >> Sweet. [77:56] >> So, it told me what it did. It created [77:58] all these new files. It modified these [77:59] existing uh files. Um and it did tell me [78:02] that I can't build because I don't have [78:04] my build set up. But we knew that was [78:06] going to be the case. [78:07] That is so cool. Okay, so right now for [78:10] slashcomands there is many and there is [78:12] a feedback slashcomand that we can uh [78:15] actually send feedback directly to the [78:17] team from here. [78:18] >> Uh you just I think I don't know what [78:21] this does. Does this open an issue? What [78:23] does it do? [78:23] >> Yeah. Yeah, it uses it uses the issues [78:25] in our um in our in our public [78:26] repository. [78:28] >> Perfect. Y [78:29] >> there's a question about can you use a [78:31] model uh can you show us how to select [78:33] the models and switch because this is [78:35] also another thing that I think we don't [78:36] talk about enough. [78:38] >> Yes. [78:39] >> So we're just going to do slashmodel and [78:41] then here are all the models. [78:43] >> Look at that. [78:44] >> I promise you the things that say hidden [78:46] aren't as um aren't as interesting as as [78:50] you might think they are. So don't don't [78:52] read into that too much. Um but uh [78:55] they're uh you know yeah you know you [78:57] just do slashmodel. can search. So if I [78:59] wanted to look for haik coup, you know, [79:01] codecs, here's all the ones that have [79:02] codecs in them. Um, if I select GPT 5.3 [79:06] codeex as an example, I can select the [79:09] reasoning effort that it uses as well. [79:10] So, you know, I want extra high. Let's [79:12] go to extra high. [79:14] >> Um, and yeah, you know, there's this [79:16] list is ever expanding. Um, we made [79:20] >> literally as they come out, they get in [79:22] this list. [79:23] >> Literally as they come out. Um, it's [79:25] it's nuts. Um but uh yeah, it's it's [79:28] it's that easy. Uh and if you're in [79:30] headless mode and you want to select the [79:31] model, um you could do copilot um-model. [79:35] Um is it going to tell me how to uh you [79:38] know, then I give the the model name? Uh [79:41] I want to use help. So you can see uh [79:44] what all of the model So here you go. [79:46] Here are all the things that you could [79:48] um [79:49] >> that you could select. Um and so yeah, [79:52] so whether you're in interactive mode or [79:54] in um non-interactive mode, you can uh [79:56] select the model that you use and it [79:58] really is one of the one of the [79:59] superpowers of this tool. [80:01] >> Amazing. Amazing. Use the best harness [80:03] and you get to use the best models like [80:06] >> Yeah. Yeah. [80:07] >> state-of-the-art and everything cost you [80:09] one premium request right now. So which [80:11] is incredible when you're thinking about [80:13] the amount of implementation and work [80:14] that these agents are doing for you. [80:16] >> It truly is. Uh so our code reviews uh [80:18] agents finished. So, we had our three [80:20] code review agents, one running uh 5.3 [80:22] Codeex, one running Opus 4.6, one [80:25] running Gemini 3. Um, and the the [80:28] reviews are in all three reviewers have [80:30] these same uh number one issues. It [80:32] looks like there was just like a [80:33] straightup bug in that implementation uh [80:34] where the punch handler was applying to [80:37] all party games and not just uh tag. So, [80:40] even in other party games um like damage [80:42] is being canceled and whatever. So, um [80:45] that's probably something we would want [80:46] to address. There's a race condition in [80:48] onstart. Um, so cool, right? You know, [80:51] now now that I had these three agents [80:53] go, um, and you can see they they did [80:54] find different things like only opus [80:56] found this race condition, only [80:58] interesting. Yeah. Found this thread [81:00] safety issue. Um, and and and you know, [81:04] that's I think part of the reason why um [81:06] I would make this a part of your [81:07] everyday workflow is having um multiple [81:10] models review your review your code. [81:12] >> Amazing. And the uh co-pilot code review [81:16] as well is part of this. You can with [81:18] the coding agent when you assign the [81:19] task that's also like baked in there. So [81:21] you're you're already getting a better [81:25] quality uh PR that you would otherwise. [81:28] >> It's true. But if you're anything like [81:30] me, you want to do some review on your [81:31] own terminal before you push it out to [81:33] the world where the rest of your team [81:34] sees it. [81:35] >> Very fair. Very fair. Um let's take this [81:38] comment from Scott. I'd love to see some [81:41] the CLI share MCP config with VS Code [81:45] and if I'm greeted even BS 2026. [81:49] >> Yeah. Yeah. So, um the the reverse is [81:52] true. So, if your repository has um um [81:56] uh VSC codemcp.json, [81:59] we will read that. Um [82:01] >> I don't know if the reverse is true. I'd [82:04] have to ask my friends on the VS Code [82:06] side. um user level configs. I believe [82:11] we do not um uh share a config at this [82:14] time. Um but uh I would invite you to go [82:17] open an issue in the public repo um or [82:19] search to see if anyone else has opened [82:20] one and thumbs it up. Uh because that's [82:22] definitely um I think a good idea. I [82:25] think the fractured nature of a lot of [82:27] these um configs across uh lots of [82:30] different agenda harnesses, that's [82:31] definitely a little bit of a bummer [82:33] right now and I would like to see uh [82:35] that improved. But uh and Scott, I'll [82:39] also give you this um bit of advice. Uh [82:42] package all your MCP servers into [82:44] plugins. That way, wherever you are, [82:45] it's easy to just install them uh in one [82:48] fell swoop. [82:49] >> Oh, that's a very good tip. That is a [82:52] very, very good tip, Ryan. This has been [82:54] so informational. I feel like I learned [82:57] a ton. You have to tell us what's next. [83:00] I have to get you to tell us something [83:01] that you haven't said to anybody. [83:04] >> Something that I haven't said to [83:05] anybody. [83:05] >> Trying to bring some a little bit of [83:07] tech journalism here. [83:08] >> Okay. Okay. Okay. [83:09] >> I know there is a lot of plans. There's [83:11] a lot of plans. When I tell you this [83:12] team is cooking, they are absolutely [83:14] cooking. Uh anything that you can share? [83:16] I mean, there is some exciting events [83:18] coming up. I know that it was just [83:19] announced that Build is happening um in [83:22] June. I mean, there'll be things cooking [83:25] before. [83:26] >> Oh, [83:26] >> and there will for sure be things [83:28] cooking for that. [83:29] >> There most certainly will. There most [83:30] certainly will. Um, I will tease you [83:33] guys and say there's been a lot of UI [83:36] improvements and a lot of improvements [83:38] to the design of the co-pilot CLI. Um, [83:40] this past month or two, you haven't even [83:44] seen 10% of it. There are some really [83:48] cool things that um the team is uh is is [83:51] is cooking on right now and I [83:56] I'm I'm very excited for what uh [83:58] visually the co-pilot CLI might look [84:00] like in uh in a few months time. [84:02] >> I love this for us especially because I [84:05] use it every day. Ryan, this has been [84:08] absolutely delightful. Thank you so much [84:09] for coming by and taking your time to [84:12] like taking us through like this your [84:13] very own personal workflow, right? Oh [84:16] yeah, we got personal. Yeah, [84:22] >> every day I feel like I see something [84:23] new. So I can't wait to see. I mean, [84:26] honestly, since September, we're in [84:28] March, friends. This is not even a year [84:30] old. [84:30] >> I know. I know. [84:32] >> It's absolutely mind-blowing. I can't [84:35] wait to see what people are are cooking. [84:37] I think and we need to do another [84:39] segment because I we kind of we covered [84:41] a ton but I'd love to talk a little bit [84:43] more about the uh the RPC the JSON [84:46] integration like being able to bring the [84:48] CLI into your own things and using the [84:50] SDK to do that. That is so powerful. Um [84:53] I don't think we even scratched the [84:55] surface of what you can do there. [84:56] >> Yeah, we'll have to follow up on that. [84:58] >> Yes. Yes. Yes. Yes. Um Ryan, where where [85:01] do you hang out? like what special [85:03] social media hell are you in that you [85:05] would like people to follow you? [85:07] >> Yeah. Uh go follow me on uh Twitterx. [85:10] I'm at ryan_heck. Um and that's going to [85:13] be the best place to uh to get in touch [85:15] with me um in real time if you want to [85:17] chat all of this. Uh but um yeah, you [85:20] might see me on on Reddit as well. I'm [85:23] Ryan_Github [85:24] I think on Ryanhack GitHub I believe is [85:26] my Reddit handle. I'm in the GitHub [85:28] copilot subreddit all the time there as [85:29] well. Um, and if you talk about the [85:31] co-pilot CLI on LinkedIn, you might see [85:33] me and your likes or in the comments as [85:35] well. [85:35] >> Yes, I forget to I always forget about [85:37] this Reddit and I know a lot of you [85:39] friends live in there. I will share the [85:42] we actually have a subreddit. They have [85:44] co-pilot subreddit and I know Ryan and [85:46] team from the CLI team, they do spend [85:48] some time there. There is also sometimes [85:50] there'll be AMAs. Um, again, this is [85:52] beautifully in the open. So, if you want [85:54] to open an issue in the repository, you [85:56] can also take a look at what's already [85:58] there. Uh, but can't wait to see what [86:01] everyone is taking, please take it for a [86:03] spin and let me know how how it went. [86:05] Let me know how you like it. Um, and if [86:07] you're a person especially who've not [86:09] tried it in in a bit, like say you try [86:12] it when it first came out. I mean, even [86:14] the installation, there were so many [86:15] things. Everything is completely [86:16] different now. Um, let me know. Let me [86:19] know how you got on. Let me know what [86:21] your experience was. I know that again [86:24] they deeply care about this being the [86:26] absolute poss best possible aging [86:28] hardness there is and offering you the [86:30] choice of the best possible frontier lab [86:32] models like anything like the sky is the [86:35] limit not even the sky on what you can [86:37] build there. So I encourage everyone to [86:39] give it a go if especially if you [86:41] haven't worked with it in a while. I'm [86:43] going to share on the chat also a link [86:45] from a recording that we did on what was [86:48] new with the coding agent because I [86:50] think that Tim also mentioned a few [86:52] things about the CLI that will be [86:53] interesting for folks. Um especially [86:55] when we're talking about a lot of the [86:57] remote sessions and being able to like [87:00] >> send a task to the cloud and then pick [87:01] it back up like so take a look at that. [87:04] Uh Ryan, come back please. [87:07] >> I would love to say the word. I'm here. [87:08] I'd love to. [87:09] >> Awesome. I'm already super selfish. I [87:11] kept you for like an hour and but way [87:14] more than I [87:16] 45 minutes. And [87:17] >> this is the most fun part of my day. I [87:19] am I am more than happy to be here. [87:20] >> Amazing. And thank you everyone so much [87:22] for being here today. Don't forget to [87:24] follow Ryan. I'm gonna find your uh ex [87:27] here. If you could drop it to me on the [87:28] private chat earlier, but now I can't [87:30] find it. [87:31] >> Yeah. [87:32] >> And preach, Scott. Preach. The CLI has [87:35] come a long way. So has Copilot [87:37] honestly, but the CLI in particular, [87:39] it's got my heart. I'm going to share [87:41] Ryan's um X. So go give him a follow and [87:46] we'll see you next. Ryan, thanks for [87:48] being here. [87:48] >> Yeah, thanks for having me. Bye folks. [87:51] >> How fun was that? I am obsessed with [87:54] this idea of the CLI teaching me how to [87:57] work with the CLI better. What? So if [88:00] you haven't tried Chronicle, try it. [88:04] This gave me it kind of brought this [88:05] idea to me. I don't know how many of you [88:07] do daily standups, but this can be your [88:09] daily standup flow at the beginning of [88:12] the day. Then at the end of the day when [88:13] you're done with whatever you were [88:15] working on, you can actually have that [88:17] short kind of interaction with the CLI [88:20] and figure out how you could have [88:21] prompted better. If there are things [88:23] like I love the advice, maybe you should [88:25] use control G. You can learn so much [88:27] just by having a conversation in natural [88:29] language. [88:31] Amazing. My mind is blown. Go follow [88:33] Ryan, please. And for those of you who [88:36] were joining from our stream on LinkedIn [88:39] or Twitch and other places, um, guess [88:42] what? We have a YouTube channel where [88:44] all of this content lives forever and [88:46] ever and ever. So, please join our [88:49] YouTube community. There's also uh [88:52] podcast that lives there and there are [88:54] community conversations, some [88:56] interesting tips that only get published [88:59] there. So, please go follow us on [89:00] YouTube. I'm like really loving to see [89:03] the growth of this YouTube account and [89:05] please help me get it. Let's push those [89:07] numbers up and then you will also see my [89:09] face in some of our recurring shows. So [89:12] if you are interested in open source and [89:14] Fridays we do an open source Friday [89:16] stream and then we also have this [89:18] shorter format but high impact checkout [89:21] segments where you get to learn about a [89:23] lot of things that especially the things [89:25] that are new and this is the first time [89:26] that we actually recorded with Ryan. [89:29] Just today we published a brand new [89:30] GitHub checkout on Agentic Workflows, [89:33] which is if this blew your mind, go [89:36] check out Agentic Workflows because it's [89:37] a great way for you to bring that [89:40] agentic powers to everything that you're [89:42] working on and do it in a way that it's [89:44] kind of like delegate. Like the task are [89:47] running, the agents are running. I'm [89:48] using it actively in some of my projects [89:50] to keep up with documentation, which [89:52] used to be a very manual task. I don't [89:54] even need to assign an issue to coding [89:55] agent because I'm running an agent [89:57] workflow that does it for me [89:58] automatically. So go check out that [90:00] YouTube please. I appreciate you all so [90:03] much. I am going to take this [90:04] recommendation to hard Ezekiel [90:07] um copilot VS Code was a bit small and [90:10] hard to read 100%. I will make sure that [90:13] we zoom in a bit next time. Please thank [90:15] you. I'm guilty of not doing you know [90:18] technically for all of you folks out [90:19] there that love dark mode. is really bad [90:21] for people to be able to see. I know I'm [90:24] just as bad though because I'm like but [90:25] my eyes. So, I will take heat into that [90:28] recommendation. Thank you all so much [90:30] for being here. Don't forget to go [90:31] follow Ryan on Twitter. Keep up with the [90:34] change log and watch our shows on [90:36] YouTube. It's a privilege to spend some [90:38] time with you today. Thank you. I'm [90:40] going to leave you with my favorite [90:41] jingle. Let's see. Where are you? And [90:44] you'll see our usual suspect here. Let [90:47] me know what you built. Let me know if [90:49] you tried the copilot CLI back in [90:51] September, October, November, even [90:53] December and you upgraded it now. What [90:56] did you think? Add me on social, please. [90:58] See you next time. [91:13] Heat. Heat. [91:23] Heat. Heat.