How I Build Agents That Work the Night Shift

How I Build Agents That Work the Night Shift ![rw-book-cover](https://img.youtube.com/vi/eqbfokhLioY/maxresdefault.jpg) 🔗 **Source:** [How I Build Agents That Work the Night Shift](https://www.youtube.com/watch?v=eqbfokhLioY) > [!note]+ Document Notes > tags: #ai-agents #automation #workflow-optimization #productivity #recurring-tasks #github #seo #system-design > author: Brian Castle > source: YouTube > reference: https://www.youtube.com/watch?v=eqbfokhLioY > date created: 2024-12-19 10:30 > > # 🏷️ Title: How I build agents that work the night shift >  ## 🧰 Action Inbox - [H1] Design recurring agent workflows using the "night shift" pattern — replaces manual prompt sessions #automation - [H2] Set up shared interface for agent-human communication — markdown files or custom APIs work #system-design - [H3] Schedule agents for specific recurring tasks like SEO audits and GitHub PR reviews — delegate routine maintenance #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 - [Builder Methods](https://buildermethods.com) - [Builder Methods Pro](https://buildermethods.com/pro) ## 📝 Summary - Brian Castle demonstrates how to create AI agents that work autonomously on recurring business tasks while you sleep - The "night shift" pattern involves three components: shared interface, scheduled agent with skills, and human review sessions - Two real examples shown: automated SEO meta tag reviews and GitHub pull request management - Key shift is from prompt-driven interaction to delegation of complete recurring workflows - System requires upfront design work but eliminates daily manual oversight of routine tasks ## 📖 Main Content The core argument is that AI agents should function as autonomous teammates rather than chat-based assistants. Instead of spending time in chat windows prompting and re-prompting, Castle advocates for designing systems where agents have recurring "shifts" and complete work independently. The night shift pattern has three essential components: 1. **Interface**: A shared source of truth (markdown files, custom apps with APIs) where both human and agent can read/write status updates 2. **Scheduled Agent**: An agent with step-by-step skills that runs automatically on a schedule (daily, weekly, etc.) 3. **Human Review**: Short, focused sessions (2-20 minutes) to review agent work and provide high-leverage feedback Castle demonstrates two production examples: an SEO agent that reviews website meta tags every two weeks and identifies optimization opportunities, and a GitHub agent that reviews open-source pull requests and makes merge/decline recommendations. Both agents produce detailed reports and handle routine maintenance that would otherwise consume significant manual time. The approach requires significant upfront system design but transforms the human role from constant oversight to strategic review and approval. Castle emphasizes this pattern works across different AI platforms (Claude, OpenAI) and can be implemented with various scheduling systems. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/eqbfokhLioY?t=0) — [H1] Night shift agents work autonomously - Key idea: Agents review websites and code contributions while creator sleeps, requiring only 10 minutes of human feedback - Evidence: "While I was sleeping last night, this agent reviewed every page on my website. It drafted SEO improvements, and it left me a report to look at this morning." #### ⏱️ [01:32](https://youtu.be/eqbfokhLioY?t=92) — [H2] Traditional AI usage keeps humans in the loop constantly - Key idea: Current AI workflow requires opening chats, prompting, waiting, and repeating in endless cycles - Evidence: "You realize you've been sitting here all afternoon staring at this chat window" #### ⏱️ [02:48](https://youtu.be/eqbfokhLioY?t=168) — [H1] Shift from prompting AI to delegating jobs to AI teammates - Key idea: Reframe AI as teammates with recurring shifts rather than tools requiring constant prompting - Evidence: "A teammate doesn't wait for you to open a chat. A teammate has a job. They show up, they do the work, and they bring you the things that actually need your attention." #### ⏱️ [03:36](https://youtu.be/eqbfokhLioY?t=216) — [H1] Night shift pattern has three core components - Key idea: Successful autonomous agents require shared interface, human review sessions, and scheduled agent execution - Evidence: Pattern includes "interface where we track status," "me, the human in the loop" for focused sessions, and "agent with a skill" running on schedule #### ⏱️ [04:58](https://youtu.be/eqbfokhLioY?t=298) — [H2] Interface serves as single source of truth - Key idea: Shared interface can be simple markdown files or complex custom apps with APIs - Evidence: "Your interface can be as simple as a markdown file with some notes and checklists" #### ⏱️ [06:45](https://youtu.be/eqbfokhLioY?t=405) — [H2] SEO meta tags require constant maintenance as businesses grow - Key idea: Active businesses constantly add pages needing optimized titles and descriptions for search visibility - Evidence: "If you have a highly active business like I do, you're probably constantly adding new pages and sections" #### ⏱️ [09:24](https://youtu.be/eqbfokhLioY?t=564) — [H3] Custom APIs enable agent automation of manual tasks - Key idea: API access allows agents to both read current state and make updates to business systems - Evidence: "Everything that I can do in the interface myself is also available in our private API that I make available only to my agents" #### ⏱️ [11:42](https://youtu.be/eqbfokhLioY?t=702) — [H2] Skills are markdown files with step-by-step agent instructions - Key idea: Reusable skills contain detailed processes for agents to follow consistently - Evidence: "A skill is essentially a markdown file...instructs the agent to review all the pages on my site" #### ⏱️ [13:18](https://youtu.be/eqbfokhLioY?t=798) — [H3] Scheduled tasks run automatically without human initiation - Key idea: Tasks dashboard enables recurring execution every 2 weeks, daily, or custom schedules - Evidence: "SEO metatags review...run that every 2 weeks on Tuesday at 2:00 a.m. in the morning" #### ⏱️ [16:24](https://youtu.be/eqbfokhLioY?t=984) — [H2] Open source maintenance creates ongoing pull request review burden - Key idea: Popular GitHub projects generate continuous contributions requiring code review and merge decisions - Evidence: "Pull requests tend to pile up every every week or so. I'll get, you know, new additions" #### ⏱️ [19:12](https://youtu.be/eqbfokhLioY?t=1152) — [H3] Agent provides detailed reasoning for pull request recommendations - Key idea: Agents analyze code changes and draft decision rationale for human review - Evidence: "The agent went through and it gave me its reasoning...it reviewed all the code and it gave me its reasoning as to why it made that recommendation" #### ⏱️ [22:36](https://youtu.be/eqbfokhLioY?t=1356) — [H2] Pattern works across different AI platforms - Key idea: Night shift approach is platform-agnostic and can be implemented on Claude, OpenAI, or other systems - Evidence: "I've built this pattern on OpenClaw and I've built it on Claude...the pattern is portable" #### ⏱️ [24:12](https://youtu.be/eqbfokhLioY?t=1452) — [H1] Identify delegation opportunities through repetition analysis - Key idea: Tasks done before that will be done again are prime candidates for agent delegation - Evidence: "Have I done this before? and will I need to do it again? If the answer to both is yes, that's a candidate to delegate to an agent" ## 📚 References - [Builder Methods](https://buildermethods.com) - [Builder Methods Pro](https://buildermethods.com/pro) - [Agent OS on GitHub](https://github.com/) - [Claude AI Platform](https://claude.ai/) ## 🗣️ Paraphrase - Castle transforms AI from reactive chat assistant to proactive autonomous teammate by designing recurring workflows that run on schedules - The system requires building shared interfaces and detailed process documentation upfront, but eliminates constant human oversight of routine tasks - Real business examples demonstrate how agents can handle SEO maintenance and code review workflows that previously consumed significant manual time ## 📰 Newsletter **Transform Your AI Workflow: From Chat Assistant to Autonomous Teammate** Many WordPress professionals spend hours each day in AI chat windows, prompting and re-prompting for tasks like SEO optimization, code reviews, and content audits. Brian Castle's "night shift" pattern offers a better approach: design AI agents that work autonomously on recurring business tasks. **The Night Shift Pattern:** - **Shared Interface**: Use markdown files or custom apps where both you and agents track progress - **Scheduled Execution**: Agents run automatically (daily, weekly) without human initiation - **Focused Review**: You spend 10-20 minutes reviewing agent work, not hours managing prompts **WordPress Applications:** - **SEO Audits**: Agents review meta titles/descriptions across all pages, fix gaps automatically - **Plugin Reviews**: Automate security scanning and update recommendations - **Content Maintenance**: Check for broken links, outdated information, accessibility issues The key shift: Ask "who can I delegate this to?" instead of "how can I do this faster?" Identify recurring tasks you've done before and will do again—those are prime candidates for agent automation. Resources: [WordPress Developer Handbook](https://developer.wordpress.org/), [WP-CLI Documentation](https://wp-cli.org/), [Accessibility Handbook](https://developer.wordpress.org/themes/accessibility/) ## ✅ Action Items #### DevRel - Developer experience: Create agent workflows to audit documentation completeness and accuracy on recurring schedules - Community management: Design agents to monitor support forums and surface high-priority issues requiring expert attention - Developer marketing: Set up agents to review content calendars and ensure technical accuracy across campaigns #### WordPress - Core: Implement agents for automated performance regression testing and security vulnerability scanning - Documentation: Deploy agents to identify outdated code examples and broken links across developer resources - Accessibility: Create recurring agent audits for WCAG compliance across WordPress.org properties - Meta: Design agents to monitor API reliability and surface degradation patterns requiring investigation ## 🧪 Hypotheses - Autonomous agent workflows reduce time spent on recurring maintenance tasks by 70-80% while improving consistency - The upfront system design investment pays back within 4-6 weeks for tasks performed more than twice monthly - Human review quality improves when focused on strategic decisions rather than routine execution oversight - Agent-generated reports with detailed reasoning enable faster and more confident human decision-making ### 📊 Methodology - Time-tracking study comparing manual task completion vs agent-assisted workflows - Quality analysis of agent recommendations vs human decisions on pull requests and SEO optimizations - Survey measurement of developer satisfaction with agent delegation vs traditional AI chat assistance ### ❓ Questions - How do you handle agent errors or edge cases that don't fit the predefined skill instructions? - What's the optimal frequency for different types of maintenance tasks when delegating to agents? - How do you maintain agent skill quality and accuracy as business requirements evolve? ### 🔮 Future Research - Integration patterns between agent workflows and existing business tools (CRM, project management, monitoring) - Best practices for agent skill versioning and A/B testing of different process approaches - Multi-agent coordination for complex workflows spanning different business domains ### 💡 Implications - Practice: Shifts developer role from tactical execution to strategic system design and review - Policy: Requires establishing clear guidelines for agent decision-making authority and escalation paths - Security: Need for careful API access controls and audit trails for automated system changes - Education: Demands new skills in workflow design, process documentation, and agent instruction writing ### 🧠 Learning Styles Assessment Reading level estimate: 9th Grade (Flesch-Kincaid) #### 👀 Visual learners - Present: Screenshots of agent dashboards, reports, and task scheduling interfaces - Improve: Workflow diagrams showing agent-human interaction loops, system architecture visuals #### 🔊 Aural learners - Present: 25-minute video with detailed verbal explanations and examples - Improve: Audio-only version for podcast consumption, transcript with speaker emphasis #### 🗣️ Verbal learners - Present: Clear terminology definitions (skills, night shift, interface), concrete examples - Improve: Glossary of agent workflow terms, comparison tables between patterns #### 🧩 Kinesthetic learners - Present: Real business examples with actual code and interfaces - Improve: Step-by-step setup guides, hands-on tutorials for building first agent workflow #### 🧠 Logical learners - Present: Three-component system breakdown, clear problem-solution structure - Improve: Decision trees for identifying agent-suitable tasks, ROI calculation frameworks #### 🧑‍🤝‍🧑 Social learners - Present: References to Builder Methods Pro community - Improve: Discussion prompts for sharing agent workflow ideas, peer review frameworks #### 🧘‍♀️ Solitary learners - Present: Individual reflection prompts about current workflows and automation opportunities - Improve: Self-assessment checklist for agent readiness, private worksheets #### 🌿 Naturalistic learners - Present: Organic business growth requiring scalable automation solutions - Improve: Ecosystem thinking about agent workflows, sustainable automation practices ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Video has clear verbal descriptions of visual interfaces - Improve: Detailed alt text for dashboard screenshots, transcript with UI element descriptions #### Keyboard navigation - Present: Web-based video player with standard controls - Improve: Transcript navigation shortcuts, timestamped section jumping #### Contrast and readability - Present: Clear text overlays in video, good contrast in dashboard screenshots - Improve: High-contrast transcript version, larger text options #### Text resizing and reflow - Present: Responsive video player - Improve: Scalable transcript layout, mobile-optimized viewing #### Media and captions - Present: Full transcript provided with timestamps - Improve: Closed captions with speaker identification, downloadable SRT file #### Forms and interactions - Present: Simple video controls - Improve: Accessible transcript search, keyboard shortcuts documentation ### 🌍 DEIB Assessment #### Cultural breadth - Present: Business-focused examples applicable across industries - Improve: International time zone considerations for "night shift," global business examples #### Inclusive language - Present: Gender-neutral "teammates" terminology, clear technical explanations - Improve: Plain-language alternatives for technical jargon, avoid sleep-centric metaphors #### Neurodiversity support - Present: Well-structured content with clear sections and timestamps - Improve: Content warnings for flashing video elements, processing time considerations #### Representation - Present: Solo presenter sharing personal business examples - Improve: Diverse use cases beyond tech entrepreneurship, accessibility of custom development approach #### Psychological safety - Present: Encouraging tone about learning new AI patterns - Improve: Acknowledge barriers to custom development, provide low-code alternatives ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear logical flow from problem (time in chat windows) to solution (autonomous agents) with supporting examples - Inductive strength: Two detailed business examples support broader pattern applicability - Causal reasoning: Demonstrates clear cause-effect between system design effort and operational efficiency gains - Analogical strength: Effective "teammate" analogy helps reframe AI interaction paradigm #### Fallacies - Survivorship bias: Only shows successful agent implementations, doesn't address failures or limitations - Appeal to novelty: Implies newer autonomous approach is inherently better without discussing tradeoffs #### Tone of voice - Overall tone: Confident and practical, with concrete examples backing claims. Maintains accessibility while demonstrating technical depth. ### 🛡️ DARVO Analysis - Deny: No defensive patterns detected - Attack: No attacks on alternative approaches or critics - Reverse victim/offender: No victim/offender dynamics present - Emotional cues and rhetoric: Positive, solution-focused presentation without manipulative elements --- ## Full Transcript [00:00] While I was sleeping last night, this [00:02] agent reviewed every page on my website. [00:04] It drafted SEO improvements, and it left [00:06] me a report to look at this morning. [00:08] Also, while I was sleeping, this other [00:10] agent reviewed every open-source [00:12] contribution that came in, and it queued [00:13] up the next action for my approval. Now, [00:16] I didn't prompt either one. I didn't [00:17] even open up a chat window. They just [00:19] showed up to work. And this morning, [00:21] I'll spend 10 minutes giving feedback, [00:23] and then they'll keep going from there. [00:24] This is what I call the night shift. And [00:26] it's a design pattern that I now use for [00:28] almost every recurring job that I [00:30] delegate to agents in my business. I [00:32] want to show you exactly how this works [00:34] because once you see it, you'll spot [00:35] dozens of places where you could use it [00:37] in your work. Now, if you're like me, [00:39] I'm sure you've become a pretty heavy [00:40] user of Claude or Chat GBT, using it as [00:43] an assistant in all sorts of tasks that [00:45] you work on. And the more you use it, [00:47] the more its memory kicks in and it [00:48] becomes an even better assistant. That's [00:50] a good start. But notice the shape of [00:52] it. Every task begins with you opening a [00:55] chat and you prompt. You wait, you [00:57] review, you prompt again. And then you [00:59] realize you've been sitting here all [01:01] afternoon staring at this chat window. [01:03] And so maybe you've taken that next [01:04] step. You've started to use skills, [01:06] which are great because you can bundle [01:07] reusable instructions so that you're not [01:09] retyping the same prompts every time. [01:11] That's progress, but the shape is still [01:13] the same. You still have to invoke the [01:15] skill. You still have to sit there while [01:17] it runs. You're still the one starting [01:19] and overseeing every job. So it feels [01:21] productive. You're using the latest [01:23] tools and you're moving faster than you [01:24] used to, but your day is still spent in [01:26] that chat window pulling the slot [01:28] machine hoping that your next response [01:30] is the one that you wanted. And if it's [01:31] not, you go in circles reprompting and [01:34] redoing the work. That was me until [01:36] pretty recently. And then I realized [01:37] something that completely changed how I [01:39] work with AI agents. The shift happened [01:41] when I started thinking of AI agents as [01:44] teammates that do jobs in my business. A [01:46] teammate doesn't wait for you to open a [01:48] chat. A teammate has a job. They show [01:50] up, they do the work, and they bring you [01:52] the things that actually need your [01:53] attention. So, I asked a different [01:55] question. Not how do I prompt AI better, [01:58] but what would it look like if an AI [02:00] agent had a recurring shift in my [02:02] business? And what if while I sleep or [02:04] work on other things, that agent's work [02:06] just gets done so that when I'm back, [02:08] there's progress waiting for me to [02:10] review. That's the pattern that I now [02:12] call the night shift. And here's how [02:13] this loop actually works. It has three [02:15] parts. The first part is the interface. [02:17] This is where we track the status of [02:19] things. It's where information gets [02:21] stored. It's where the agent and I [02:23] communicate, but not like a live chat. [02:25] Your interface can be as simple as a [02:27] markdown file with some notes and [02:28] checklists. Or for more complex [02:30] processes, it could make sense to custom [02:33] build an app with a user interface that [02:35] you can use and an API that your agents [02:37] can use. The key here is to establish [02:39] one source of truth that you and your [02:41] agent can read and write to. By the way, [02:44] that custom app route these days, it's [02:46] more doable than you think, even if you [02:48] don't have coding experience. I have [02:49] some free open- source templates and [02:51] skills to help get you started. And if [02:53] you're a Builder Methods Pro member, I [02:55] created a course on becoming a builder [02:57] with AI. The second part is me, the [02:59] human in the loop, but not in the way [03:01] that you might expect. I'm not in there [03:03] all day. I drop in for short, focused [03:05] sessions, sometimes 2 minutes, sometimes [03:07] 20. I review the agents work. I leave a [03:10] comment or I approve something or I [03:11] check a box. and then I close it and I [03:14] get back to my other work. The third [03:15] part is an agent with a skill, which is [03:17] a step-by-step process. I set up my [03:20] agents to run each skill on a recurring [03:22] schedule. So, every time it runs, the [03:24] agent picks up wherever we left off. It [03:25] acts on my latest feedback. It pushes [03:28] some work forward and surfaces some new [03:30] items for me to review next time. I call [03:32] this the night shift because most of my [03:34] agents run overnight while I'm sleeping. [03:36] But they don't have to. The schedule is [03:38] whatever fits the work. Daily, weekly, [03:40] every other Tuesday. The point is that [03:43] it's recurring and it's automatic. Now, [03:45] the real work in all of this isn't the [03:47] agent doing its job. It's in you [03:49] designing the system upfront. The [03:51] interface, the skill, the process the [03:53] agent will follow. That's the muscle [03:55] that we need to build as operators in [03:57] this age of AI. But that's what we're [03:59] building here. A system with three [04:01] parts. A shared interface, an agent on a [04:04] schedule, and you giving high leverage [04:06] input. Now, I'm going to show you two [04:07] real examples running in my business [04:09] right now. But first, if you're new [04:10] here, I'm Brian Castle. I help builders [04:13] stay ahead of the curve with AI. And [04:15] every Friday, I send out my builder [04:17] briefing. It's a free fiveminute read on [04:20] where the shifts are happening as we [04:21] build with AI. No hype, just practical [04:24] stuff from the trenches. You can get [04:25] yours by going to buildermethods.com. [04:28] All right, the first real example that [04:29] I'll show you, this is a real agent [04:31] process that's running in my business. [04:33] It has to do with SEO, something that [04:35] basically every business has had to [04:37] think about. And these days, you know, [04:39] we're thinking about AI chatbot SEO, [04:41] which a lot of the same best practices [04:43] apply. When it comes to SEO, really one [04:46] of the most important lowhanging fruit [04:49] things that you can do for your website, [04:52] is to make sure that you have optimal [04:55] metatitle tags. That's the text that [04:58] shows up on the browser tab and [05:01] metadescription tags on every one of [05:03] your pages so that they are accurate and [05:06] they describe exactly what's on the page [05:08] and they help someone who is searching [05:10] in Google or asking chat GPT about your [05:12] website that's what's going to come up. [05:14] So, for example, like in Google, if I [05:16] Google builder methods, of course, we're [05:19] at the top here, but you know, the the [05:20] meta title tag is what shows up here, [05:23] and the meta description tag is what [05:25] shows up here. And so, it's really [05:27] important that as you add lots of new [05:29] pages to your website, you know, each [05:32] one of these pages needs to have [05:34] descriptive title and meta tag that [05:38] accurately describes what is actually on [05:41] that page. So that and you know if you [05:43] have a highly active business like I do, [05:46] you know, you're probably constantly [05:47] adding new pages and sections like I'm [05:50] always adding new videos and courses and [05:54] new features, new new build kits. All of [05:57] these need to have descriptive titles [05:59] and metatags. And it's really easy for [06:01] these things to drift and sometimes you [06:04] forget to add them. And sometimes you [06:06] you you know you need to go through and [06:08] make sure that you plug these holes [06:10] because again it's that lowhanging fruit [06:11] that if you forget about it or if you [06:14] let them fall by the wayside that can [06:16] actually be really damaging for your [06:18] business to show up in Google and to [06:20] show up in chat GPT and and other chat [06:23] bots. So historically that would have [06:25] been a task that either I would do once [06:28] in a while when I remember to do it or [06:30] maybe hire a marketing person or an [06:32] agency to take care of that sort of [06:34] thing. But these days we can automate [06:36] and delegate that to an AI agent. Let me [06:39] show you how I set it up. So first of [06:40] all, how do we actually manage the SEO [06:43] title and meta tags in my website? So my [06:47] site for builder methods, not [06:48] surprisingly, it's a customuilt [06:50] application and I have a custom backend [06:53] area where I can manage everything [06:55] including this customuilt SEO dashboard. [06:58] So, if you have like a like a WordPress [07:00] site or or or a CMS of some kind, you [07:02] probably have some version of this. Mine [07:05] is customuilt where every page I can go [07:08] in and actually manage the titles and [07:10] meta tags and I can go through all of [07:12] our collections of videos in the library [07:15] and all the articles and all the build [07:17] kits and everything. And I can make sure [07:19] that that the images and the titles and [07:21] descriptions are all correct. And, you [07:23] know, I can customize them if I need to, [07:25] right? If I need to fix anything up. So, [07:27] it's all managed here, but that's for me [07:30] to manage it manually, and I want to [07:33] actually delegate this and make sure [07:35] that it's always taken care of and and [07:36] that it's always checked on. So, in [07:39] addition to being able to manage it here [07:41] in the user interface, I also have an [07:43] API to access all of this stuff. So [07:47] everything that I can do in the [07:49] interface myself is also available in [07:53] our private API that I make available [07:56] only to my agents so that my agents can [07:59] go in and they can check out the SEO [08:01] titles for all of my pages. They can go [08:04] make updates to the titles and meta tags [08:06] if they need to. Okay, so we have the [08:09] source of truth, right? We have the the [08:11] management layer here and we have the [08:14] API layer so that agents could access [08:16] this. Then I have an agent that runs a [08:20] skill that's in my collection of skills. [08:22] This is like in my private collection. [08:24] So a skill is essentially a markdown [08:27] file. Uh I have mine kind of broken up [08:30] into a couple of like phases that that [08:32] it follows. But essentially, this [08:34] markdown file instructs the agent to [08:38] review all the pages on my site, all the [08:41] titles and metas, and it follows some [08:43] rules and some instructions that I've [08:46] laid out in this in the skills [08:48] instructions or step-by-step process in [08:50] here to basically identify if there are [08:53] any gaps or or any of these titles and [08:55] metas are suboptimal. So, if it does [08:58] find those, then it will use the API to [09:01] fix those things. But I don't want it to [09:04] just do the work and be invisible and [09:06] maybe make some mistakes and things like [09:08] that. So, so I also instruct it to write [09:11] up a report and send it to me. So, when [09:13] is my agent actually going to do this? [09:15] Well, I dispatch that task to my agent [09:17] using my custom tasks dashboard. This is [09:21] another customuilt app. I have a guide [09:24] for how to build your own version of [09:25] this inside builder methods pro. This is [09:27] where I dispatch tasks to all my [09:29] different agents. Currently, these are [09:31] actually all running on cloud code on my [09:35] dedicated Mac Mini. I used to have this [09:38] running on OpenClaw and I recently [09:39] migrated my uh my agents over to Cloud [09:42] Code uh mostly because I just want to [09:45] take advantage of my Claude Max plan. [09:47] Anyway, that's a side discussion for [09:48] another time. So, this admin bot has a [09:51] bunch of things that it does on a [09:52] recurring basis, but one of the things [09:54] down here says SEO metatags review. And [09:58] it's going to run that every 2 weeks on [10:01] Tuesday at 2:00 a.m. in the morning. I'm [10:04] hopefully sleeping at that time. And [10:06] I've set it up so that all this task [10:08] does is it tells it to read my skill, [10:11] which is what I put together here, every [10:13] time it runs. And then when it does [10:15] that, my agent is going to complete its [10:17] work. This was my daily report it sent [10:20] me. And then a few days ago, it sent me [10:22] this SEO review. And I can click the [10:26] link that it gave to me. And this is a [10:28] markdown file that my agent created and [10:31] saved to my files. And here I'm viewing [10:34] it. By the way, this little brainown app [10:36] that I call it, this is a markdown file [10:39] editor and reader, another custom tool [10:41] that I that I built. And I have a build [10:43] kit in Builder Methods Pro to let you [10:46] follow along and build your own version [10:48] of this as well. It's really handy for, [10:50] you know, editing and and managing [10:51] markdown files, especially when you have [10:54] agents passing around links to the [10:56] markdown files that they work on. So, [10:58] uh, really helpful. Anyway, this is the [11:01] SEO report that I'll receive from from [11:03] my agent every time it goes through and [11:06] it finds any fixes that it needs to make [11:08] to my titles and metas. So in this [11:10] particular case which ran yesterday, it [11:12] found two issues in all the pages on the [11:14] site. So it found those uh there were [11:16] two sort of like generic non-escriptive [11:19] titles and it went ahead and it retitled [11:21] those to be accurate about what the page [11:24] is all about. So just plugging those [11:26] holes really important for the health of [11:28] the SEO and GEO, whatever they call that [11:32] for my builder method site. And by the [11:33] way, that one I don't typically need to [11:36] actually take any action. I usually just [11:38] review what's done here. But if for [11:40] whatever reason, if it was like a really [11:41] important page and it just needs my [11:43] approval, then we would use these [11:45] markdown check boxes for me to check it [11:48] off. And then the next time the agent [11:50] runs, it will see what I what I left [11:53] here, either the checkbox or any notes, [11:55] and then it will take the the necessary [11:57] action. All right. So here's another [11:58] real example of an active agent process [12:01] that is happening in my business, and [12:03] I'm using the night shift model for [12:05] this. So another thing that I do through [12:07] builder methods is I release some free [12:10] open-source tools. So these are free [12:12] tools for builders. Some of them are [12:14] like agent skills, some of them are like [12:17] frameworks and things, but they mostly [12:19] live on GitHub. So a popular one that I [12:21] released last year was called agent OS. [12:24] Another one was called design OS. And [12:27] just recently uh I launched one called [12:29] PRD creator. So I'm going to be doing a [12:32] video on that in a couple days. So stay [12:34] tuned about that. These are all free and [12:35] open source. And what that means is they [12:38] tend to live on a public GitHub [12:41] repository. So when you have a public [12:43] free open- source tool on GitHub and [12:46] once it gets kind of popular, you start [12:48] to receive what are known as pull [12:50] requests. So pull requests are like [12:52] contributions from other people and they [12:55] might be bug fixes. Sometimes they're [12:57] feature additions to to the tool or the [13:00] the code base, if you will. Now, as you [13:02] can see, my pull requests for this tool [13:04] are clean. There are no open pull [13:07] requests, which is nice because it's not [13:08] usually the case. And so, what tends to [13:10] happen is pull I'm going to look at the [13:12] the closed or the completed pull [13:15] requests. These things tend to pile up [13:17] every every week or so. I'll get, you [13:19] know, new additions, whether they're bug [13:21] fixes or feature requests from different [13:23] people. And so that can be really [13:24] timeconuming for me to go in and review [13:28] every single pull request and really [13:30] inspect the code that the person is [13:33] proposing that they change or that that [13:36] they're proposing to add to my project. [13:38] So as you can imagine it's it's really [13:40] timeconuming and tedious to go through [13:43] and personally review every single thing [13:46] and then make a decision as to yes I [13:48] want to accept this request and merge it [13:51] into my codebase or yes but I need to [13:54] clean up some things before I merge it [13:56] or I could decline and close the the [13:59] pull request and and just say like it's [14:02] not something that I'm want to merge [14:03] into my open source tool at this time. [14:06] So, that's a really common thing for [14:08] anyone who who offers free open-source [14:11] tools on GitHub. But, you know, I I [14:13] don't want these pull requests just [14:14] piling up. And, uh, some of them really [14:16] are good ideas that would really improve [14:18] product. And so, what I did was I set up [14:21] an agent skill. And I'll go back to my [14:23] tasks dashboard here. My devbot, this is [14:26] the one that I assign like development [14:28] related things to down here. I've got [14:30] one called review PRs, pull requests on [14:33] my agent OS project. It's currently [14:35] running every Wednesday. I'll probably [14:37] start to increase the frequency of that. [14:39] And it's going to run the public PR [14:41] review skill. So, I have that mapped out [14:45] here. This is the skill. It's got all [14:47] sorts of instructions. It's got some [14:49] notes and rules to help it make [14:52] decisions the way that I would make [14:54] decisions when it comes to reviewing and [14:56] deciding whether a poll request is [14:58] something that I do want to merge or I [15:00] want to decline. And either way, we're [15:02] going to want to post a comment to thank [15:04] the contributor for their work. So when [15:06] my agent runs this skill on a recurring [15:08] basis, it's going to check the GitHub [15:11] repo and find any pull requests that [15:13] haven't been reviewed yet. And then it's [15:15] going to go ahead and make a decision or [15:17] make a recommendation, if you will. And [15:19] and so this is one where I actually do [15:21] want to have a little bit more hands-on [15:23] input to give the final say as to yes, [15:26] go ahead and and merge that or okay, [15:29] that's a good recommendation to close [15:30] that, but I still don't want to spend [15:32] the time personally reviewing every [15:34] single line, right? So I follow the same [15:37] pattern. The agent runs that every [15:39] couple of days and then it sends me a [15:41] message on Telegram with a link to a [15:44] report. Once again, that report is a [15:46] markdown file. I can click the link. I [15:48] open it in my markdown reader here. And [15:51] here is the report that I received a [15:53] couple of days ago. And we actually had [15:55] quite a few pull requests that had piled [15:57] up. And so the agent went through and it [16:00] listed each one and it made a [16:03] recommendation. So let's see, these top [16:05] two were actually merged. The these were [16:08] good contributions that actually fixed [16:10] small issues in in the tool. The agent [16:12] went through and it gave me its [16:14] reasoning. like it it reviewed all the [16:16] code and it gave me its reasoning as to [16:18] why it made that recommendation. Now I'm [16:20] viewing this right now in the state [16:22] after the action has already been taken. [16:24] But when I first receive this report [16:27] instead of seeing instead of it already [16:29] being done, the agent actually gives me [16:32] these check boxes where I have some [16:34] choices. All I need to do is is click a [16:37] checkbox, right? And so on these two, I [16:39] took the recommendation from the agent. [16:41] It sounded correct to me. So I I checked [16:44] this like yes merge with a comment, [16:46] right? And this has a link that goes [16:48] directly to that pull request and we can [16:50] see that the agent then the next time it [16:52] ran it saw that I that I gave the [16:54] go-ahad the green light to go ahead and [16:56] merge it. The agent did that and the [16:58] agent also went in and posted this. Now [17:00] it looks like it's actually coming from [17:02] me as a comment, but the reality is this [17:04] was actually written by the agent. it [17:07] merged. Thanks for the fix. And it [17:09] actually tagged the contributor and then [17:10] they they iconed that. You know, [17:12] sometimes the agent actually puts in a [17:14] more descriptive comment. These aren't [17:16] just template boiler plates. You know, [17:18] now the rest of these were sort of [17:19] duplicates or they weren't things that I [17:21] wanted to merge. So, the the agent made [17:23] the right decision, the right [17:25] recommendation to go ahead and close [17:28] those requests. And it's even suggesting [17:30] what the comment should be. So, the [17:32] agent actually drafted up this pretty [17:34] descriptive comment. and I can check [17:36] that out. So, we closed this one and [17:39] here is the message that was commented [17:42] sort of from me, but this was actually [17:44] drafted by my agent. So, that's just, [17:47] you know, one of many examples and a [17:50] growing list of examples of tasks that [17:52] I'm adding to my agents to-do list to to [17:56] do on a recurring basis. You know, just [17:58] really plugging all the holes in my [18:00] business. Actually, I just added another [18:02] one the other day. This is for keeping [18:04] my email marketing sequences that I have [18:06] lined up in kit. I want to make sure [18:08] that those stay up to date. So, I have [18:10] an agent sequence that runs every couple [18:12] of weeks, checks all my emails, you [18:14] know, ensures that we're not missing any [18:16] information and any new products or new [18:19] things that I want to feature are [18:21] actually getting featured in my email [18:23] newsletters or in my email sequences. [18:26] So, you know, that's another thing that [18:27] I don't want to drift and forget about. [18:29] So, I put an agent on it using the night [18:31] shift model. Now, I bet you have some [18:33] questions, so leave those in the [18:34] comments and I'll answer as many as I [18:36] can, but I bet I can guess two of them. [18:38] First, which agent platform can you use [18:40] the night shift pattern on? Honestly, [18:42] any of them work. I've built this [18:44] pattern on OpenClaw and I've built it on [18:46] Claude. I've shown both on this channel [18:48] and these days I'm building mostly on [18:50] Claude because right now I like their [18:52] models the best, but the pattern is [18:53] portable. Pick whichever ecosystem [18:55] you're already comfortable in and you [18:57] can build with the night shift pattern [18:59] on it. The second question you're [19:01] probably wondering about is how do you [19:02] actually schedule recurring tasks for [19:04] agents? Well, you've got a few options [19:06] here depending on your setup. Now, I [19:08] built my own tasks dashboard for [19:10] dispatching tasks to my agents. It lets [19:12] me set up schedules, attach agents [19:14] skills, associate tasks with different [19:17] agents, and log their activity all in [19:19] one place. And if you want to build the [19:21] same, I released a build kit for it [19:23] inside Builder Methods Pro. But you [19:25] don't need to build your own dashboard [19:27] to get started. Claude Co-work has a [19:29] built-in scheduled tasks feature and [19:31] OpenClaw has a cron system. So you can [19:34] try those out and see what fits. The [19:36] point is since the tooling around AI and [19:38] agents is changing so fast right now, [19:41] it's more important for us to focus on [19:42] our patterns like the night shift model [19:44] that I shared here so that we can adapt [19:46] into any platform. Now, here's a [19:49] question I want you to be asking when [19:50] you're sitting down to work on something [19:52] today. Ask yourself two things. Have I [19:55] done this before? and will I need to do [19:57] it again? If the answer to both is yes, [19:59] that's a candidate to delegate to an [20:01] agent using this night shift model. You [20:04] know, most of us never stop to notice [20:05] this. We just keep doing the work. The [20:08] same code reviews, the same customer [20:10] follow-ups, the same reports, the same [20:12] maintenance. We're so used to doing it [20:14] that we don't see it as something that [20:15] an agent could do for us. So, stop [20:18] asking how can I do this faster? Start [20:20] asking who can I delegate this to? And [20:23] in this era more and more that someone [20:25] can be an agent. If this is the kind of [20:27] thing that you want more of practical [20:29] patterns and systems for becoming a more [20:31] confident builder with agents, then hit [20:33] subscribe so you don't miss my next [20:35] video on the channel. And if you're [20:36] ready to go deeper with me and join [20:38] hundreds of other builders in our [20:39] community, check out Builder Methods [20:41] Pro. That's where I publish my full [20:43] courses on building with agents and AI. [20:46] We've got build kits to help get you [20:48] started and live Q&A sessions happening [20:50] monthly. Join us at [20:51] buildermethods.com/pro. [20:54] Let's keep building.  📄 **Full text:** ![[Full Document Contents/Articles/How I Build Agents That Work the Night Shift.md|How I Build Agents That Work the Night Shift by YouTube]] ## Highlights - that you might expect. I'm not in there all day. I drop in for short, focused sessions, sometimes 2 minutes, sometimes 20. I review the agents work. I leave a comment or I approve something or I check a box. and then I close it and I get back to my other work. The third part is an agent with a skill, which is a step-by-step process. I set up my agents to run each skill on a recurring schedule. So, every time it runs, the agent picks up wherever we left off. It acts on my latest feedback. It pushes some work forward and surfaces some new ([View Highlight](https://read.readwise.io/read/01krnx2dw0gw13wywx862frg46)) - Tags: 👻 ai highlighted - Tuesday at 2:00 a.m. in the morning. I'm hopefully sleeping at that time. And I've set it up so that all this task does is it tells it to read my skill, which is what I put together here, every time it runs. And then when it does that, my agent is going to complete its work. This was my daily report it sent me. And then a few days ago, it sent me this SEO review. And I can click the link that it gave to me. And this is a markdown file that my agent created and ([View Highlight](https://read.readwise.io/read/01krnx2dx01t5avzdsa8kjpppn)) - Tags: 👻 ai highlighted - probably wondering about is how do you actually schedule recurring tasks for agents? Well, you've got a few options here depending on your setup. Now, I built my own tasks dashboard for dispatching tasks to my agents. It lets me set up schedules, attach agents skills, associate tasks with different agents, and log their activity all in one place. And if you want to build the same, I released a build kit for it inside Builder Methods Pro. But you don't need to build your own dashboard to get started. Claude Co-work has a built-in scheduled tasks feature and ([View Highlight](https://read.readwise.io/read/01krnx2dz7fav0qq6s9a7dxkt1)) - Tags: 👻 ai highlighted