Making Opinionated AI Tooling Decisions With Nimbalyst's Greg Hinkle  🔗 **Source:** [Making Opinionated AI Tooling Decisions With Nimbalyst's Greg Hinkle](https://share.snipd.com/episode/f629e626-2bee-4b45-b846-8d27e009ca65) ## Highlights - **Multitasking Requires Higher Bandwidth Interfaces** - Managing many AI agents increases user bandwidth needs rather than just adding tabs. - Greg Hinkle says Nimblest creates a single pane with tools so users can multitask without losing context and control. Transcript: Greg Hinkle Yeah, we're trying to thread that needle of changing the way the experience is happening for a user by just giving them a different interface, focusing on what the user needs to see as They're managing many agents, doing many jobs. So it's really about increasing the bandwidth for the user so that they can keep control of the chaos. Scott Hanselman The chaos is true. Like I was talking with Mark Wysenovich, my buddy at work about what's the right number of agents that I can manage comfortably. And he and I feel for us, it's three to five, but then I've got guys at work and gals at work that are doing 10, 20 things, and they have squads of squads of agents and things like that. And I'm finding that the context switchings is too much. The cognitive load, for me, running all these basically terminal tabs, that's as far as I've gone, is terminal tabs. And it's hitting the wall pretty quickly. Greg Hinkle Yeah, that's exactly right. We have this opinion that keeping track of everything goes beyond just having tabs. Tabs are just too short-lived. And a lot of, you know, it's funny, it's sort of like a hurry up and wait where you come up with this great idea, but then you have to wait for the agent to go do the work or go do the research. And so you're not getting the best bandwidth if you're not multitasking. ([Time 0:01:30](https://share.snipd.com/snip/b9f2ecc1-ea66-4d30-8383-53ad2852e6f2)) - **Prep Agents Then Enter Focused Review Sessions** - Do prep work before switching into a focused session so you can evaluate agent results without frequent context switches. - Greg recommends having agents do research, then returning to spend 20–30 minutes testing, reviewing, and discussing outputs. Transcript: Greg Hinkle Exactly. It's the context switching that's the pain. So what you're really doing is trying to do all the prep work so that when you do context switch into a particular task, you can really focus and spend that 20, 30 minutes doing nothing But thinking about what did the agent come up with? Is it the right thing? Testing it, using it, looking at the code, having another agent review the code, and really talking it over with that agent. ([Time 0:03:56](https://share.snipd.com/snip/9826612c-8bd9-4cef-8221-010f452d14f1)) - **Built In Visual Tools Improve Agent Collaboration** - Embedding familiar visual tools directly into the workspace improves communication with agents and humans. - Nimblest integrates Excalidraw, ERDs, and mockups so sessions retain context across design, data model, and code. Transcript: Scott Hanselman Symbolist does well, is you're bringing some skills and some work along that is already in our lives, like Excalibur and different mockups and tools like that. And you're bringing them out of the box because I usually end up telling the agent, oh, make me some ASCII art mockup. And I get kind of a weak sauce ASCII art with a non-aligned right margin and everything like that. But you've got skills and commands built in for stuff like that. You're actually acknowledging that these tools exist and then baking them into the flow directly. That's right. Greg Hinkle We think if you have a single pane of glass that is all of the tools that you need to work with an agent to be able to build, to build product, to build technology. If all those tools are in the same place, then it's not an agent making MCP calls all over the place and you're using different interfaces. You can work on the mock-up and the design of something. You can work on the data model with the Prisma graphic design. ([Time 0:04:33](https://share.snipd.com/snip/b98beefe-8929-44ed-b69e-990d50e70293)) - **Use Collaborative Local Docs For Durable Plans** - Try local real-time collaborative docs for planning inside the workspace instead of scattering specs across external trackers. - Greg describes a local Google Doc–style markdown editor checked into Git or shared collaboratively in Nimblest. Transcript: Scott Hanselman Now, the planning stuff, you've built in a little SQL database. You've got plannings, decisions, bugs, tasks. I noticed that that is being checked in. You're versioning that. You're plugging that into the flow as opposed to using GitHub or Jira. Is that a conscious choice as well? Greg Hinkle It's one of those things we're experimenting with right now. Some of our files, we do end up checking into Git as just the mechanism to share. But we're also experimenting with basically collaborative document sharing. So real-time markdown editing. It's essentially, you know, a local Google Doc, you know, sitting inside of NimbleList. ([Time 0:06:42](https://share.snipd.com/snip/5940ce50-9e2b-4da1-907b-1088b9104b1f)) - **Show Files And Diffs To Preserve Engineering Craft** - Exposing files and diffs keeps engineering craft visible and preserves work artifacts. - Greg prefers local-first, open formats so designers and engineers can access full code and docs, not thin cloud views. Transcript: Scott Hanselman Yeah. Now, hopefully, this is a compliment. Don't take this wrong. But, like, it doesn't feel like vibes. It feels like a software development lifecycle tool. And the reason I say that is that you put the file viewer up front. You're not trying to hide files and diffs from me, which I'm finding to be somewhat unique in this space. So many AI tools are hiding my code from me. Greg Hinkle Nothing against those tools. They're fantastic for the jobs that they're trying to enable. But if the end result is great software, and it's the entire product, including all the work that you put into designs, that's all of your work. And we just think that open formats that are on disk, not sitting in some database tool or out in the cloud, or you get sort of just a thin view of it that's like, here's the running view, But to get to the code is challenging. ([Time 0:07:27](https://share.snipd.com/snip/804bc27f-a040-489a-b8c6-bf479cb2d512)) - **Mix Cloud And Local Models Based On Task Needs** - Try plugging in multiple model providers and local LLMs depending on task and context size. - Greg notes coding needs large context so cloud models often perform better, but local models are improving weekly. Transcript: Scott Hanselman Now you've got Claude, you've got OpenAI, you've got pluggable agent providers. Also notably, you have my friends at LM Studio. I love those guys. So you can connect to local LLMs. Have you had good experience with that? I feel like I need a really beefy machine to get valuable insights. Greg Hinkle I think that's true today. It depends on the tasks. Some people are using NimbleList not to just software code, but to, you know, just do prototyping or even other business tasks. And I think they're finding more success. The coding tasks with just the size of the context needed is still a challenge, but you also see advancements every week. And I think the models are going to get smaller and more efficient. ([Time 0:08:35](https://share.snipd.com/snip/6ee4894c-b75a-4f2b-8ae1-25b8dd590cb9)) - **Pointing And Drawing Reduce Instruction Friction** - Pointing, drawing and linking UI elements drastically reduce instruction friction to agents. - Nimblest's extension APIs let you click ERD tables or circle mockup regions so agents get exact targets without verbose text. Transcript: Greg Hinkle That's right. And we're experimenting with more of the kind of collaborative agent task management. So the agent can create the tasks on the Kanban and manage the flow of the sessions from planning to implementing to user validation. And all of that's customizable by the user by updating their agent files and saying, hey, when Nimblist, when you create a plan and we're going through these steps, here's the columns I want to use and here's when I want you to move it to ready for development or validating. Scott Hanselman And you can let the agent help you organize the the chaos and and it's one of the chaotic things that i find myself doing a lot is is markups where i'll have my agent on one side and then i'll Have a playwright or a markdown file or some kind of mock-up or some live html but with nimblest and help me understand how do this, you can circle the screen, say, put something there, And then mention it, like, put this here, or put it in the circle area. So you're closing a loop that I didn't realize needed to be closed? Because I usually find myself trying to describe the thing, like, use this CSS ID when I really just want to point at it and go that one. Greg Hinkle That's right. And we have, you know, within our extension system, there's this API to do a deeper level of integration with the agents so that, you know, when you're in the ERD diagram, you can click A particular, you know, table and say, these are the changes I want to make to this. And you don't have to tell it what table. It's getting that as a link. Or in a mock-up, like you said, drawing. And so we're just trying to improve that bandwidth between the user and the agent. And I think things like drawing or pointing, you can make it a lot simpler for a user to instruct an agent. Because it's really about how much control can the user deliver to the agent and how much typing. ([Time 0:09:57](https://share.snipd.com/snip/d95855ef-11e9-40e3-bf1d-00b2a1ee195c)) - **Walking The Yard Controlling Coding Agents By Voice** - Greg built a voice agent that controls coding agents so he can manage them while walking the yard. - He says it fires prompts, wakes him with notifications, and enables high-speed multi-agent control through headphones. Transcript: Greg Hinkle Yeah, I've had some moments of absolute pure delight in the sort of interactive voice. You know, it's still in testing, but I basically have a voice agent that can control a coding agent. It can control a whole set of coding agents, and I'm talking to it, and it's firing off prompts, and it's just a much higher rate of speed, and it can wake up and notify me. And I can be walking around the yard having a conversation controlling multiple agents all through my headphones. ([Time 0:13:23](https://share.snipd.com/snip/35947138-0c92-4731-b18e-c2f2c1571647)) - **Agent Harnesses Enable Goal Seeking And Iteration** - Let agents run with harnesses that expose logs, DB, and DOM so they can iterate and test autonomously. - Greg explains Nimblest's harness allows agents to read logs, query DB, take screenshots, and loop until goals complete. Transcript: Scott Hanselman You mentioned that you're using Nimbleist to write Nimblist. And I was always fascinated when I was talking to like Anders Heilsberg on, we made C-sharp and it's like, now we've got the compiler building the language, the compiler builds itself Now. How, what was that like? When did you get to the point where like Nimblist could basically talk to itself and make its own self? Because I assume you started, you didn't vibe the thing, but now you're in this line of loop where it's building its own body. Greg Hinkle Yeah, it was pretty quickly that we got to the point where the coding happened more in Nimblis than anywhere else. And I still have other IDs that I use and whatnot, but people talk about the harness. And a lot of what I've done with Nimblist over time is create more and more control for the agent with Nimblist. So as the agent is doing coding on Nimblist, it can read the log files live, it can query the database live, it can read the settings live, it navigate the UI and take screenshots and look At the DOM. And so now I'm trying to get it to control and optimize the layout of something. And I can say, here's what I want it to look like. And now it can basically loop until it's done. ([Time 0:16:55](https://share.snipd.com/snip/c9bf2a24-6180-4cd0-9a69-e59289ba0a37)) - **Chain Models For Plan Implement Review Cycles** - Use multiple models in a review cycle where one model plans, another implements, and a third reviews. - Greg calls this grouping 'work streams' and often has Claude plan, Codex implement, and another model review the code. Transcript: Scott Hanselman It. So you've got both Codex and Cloud Code today. It's all pluggable. So you're an AI native workspace built on top of each of them. Do you bounce between them? Because I use GitHub Copilot, and I feel like Claude is fine, it's great, Opus is good, but if I only had Opus available, I would feel frustrated. And I feel the same about Codex. So I like Copilot because I can bounce around between models. I'm assuming you did the same thing, and you're switching models as you do your job. Greg Hinkle That's right. Yeah. Every day I'm, I am having one agent review the work of the other agent. I find that extremely valuable. And so we have this concept in Nimblist called work streams, where I can kind of group sessions. And I do that frequently. I make a Claude Code session and then have Codex review the plan that Claude Code came up with it. You should find some errors. Then I'll have Cloud Code implement. I'll have Codex review the implementation. It's a cycle. ([Time 0:19:10](https://share.snipd.com/snip/d1072243-7f9c-49c2-b39d-d655fa98dce3)) - **Plans As Durable Why Documents Not Complete Specs** - Plans should document motivation and important constraints, not attempt to fully specify executable instructions. - Greg shifted to using plans as durable anchors that explain why and what matters, while code sculpting handles specifics. Transcript: Greg Hinkle The models have really advanced in the past year. A year ago, yes, you really needed to spend time on a plan if you were going to get valuable work out of it. I think what has happened is I've shifted. I still do plans for just about everything. But what I've focused on is making sure that the plan isn't the instruction to the agent. It's the documentation of what I want. Because I like being able to come back to it. I like being able to reference a second session to that plan and say, is this correct? Did you forget anything? This is how I want you to solve this agent. It's about here's the motivation behind it. And here are the important aspects that you can't miss. ([Time 0:25:16](https://share.snipd.com/snip/64f76eff-e864-4cce-98ae-25b222547fc0)) - **Stay Connected To Agents With A Mobile Tunnel** - Use a mobile tunneled client to stay connected and steer agents away from your desk. - Nimblest's iOS app tunnels via secure websockets so Greg can get notifications, answer prompts, and start research while walking or making lunch. Transcript: Greg Hinkle Yeah, we just wanted a you know kind of seamless native experience so we built um ios first android coming but um you know it's it's really just nimblest in a in a phone experience that's Just you know a wind a glass into back to your machine um we're doing it in encryption tunneling um just to make it seamless. Scott Hanselman You're using secure web sockets through the NimbleList servers, but it's ultimately tunneling. Greg Hinkle Yeah, exactly. And just giving you that high bandwidth ability to answer a question to the agent or to get notified and say, oh, it's done, so I can be making lunch and get notified. I find that really valuable. Sometimes I'll go for walks out in nature, and I'll be able to have a little bit of a voice conversation, kick off a few researches. I'm not necessarily coding at those times, but I'll be like, here's an idea. What about this? Go do research on that. ([Time 0:28:16](https://share.snipd.com/snip/e2d15b63-9632-43d1-a594-949f5f10e12d))