{"slug": "building-with-grok-bot-in-the-outer-loop", "title": "Building with Grok Bot in the Outer Loop", "summary": "Developer and blogger boundsj built a system combining T3 Code, his open-source T3 Fleet Gateway MCP wrapper, and xAI's Grok Bot to delegate management of multiple coding-agent threads across several machines, reporting that his own capacity to manage agent threads is the primary bottleneck on how much he can build. Grok Bot, which he describes as a mid-tier model running in an OpenClaw-style outer control loop with timed routines, skills and memories, prompts Codex and Claude Code agents through the gateway and escalates only decisions he defines as critical. He says he handles one agent thread well and perhaps up to five before things break down, and that agents outperform him at prompting coding-agent teams, merging code and releasing.", "body_md": "Over the last several weeks I’ve continued to push myself, all of my computers, and all of my [AI subscriptions](https://www.bounds.dev/posts/environment-update-a-t3-code-tokenmaxing-constellation/) to see how much I can build and how quickly I can ship it. I’m not solving any of the world’s major problems. So far I’ve just built things like a news reader, a [game](https://www.bounds.dev/posts/rescue82-making-a-choplifter-style-game-for-apple-vision-pro/), and a quantitative research desk monitoring Kalshi data.\n\nI’ve also made a few tools, like [t3-fleet-gateway](https://github.com/boundsj/t3-fleet-gateway), that have really helped me leverage the tech and accelerate way beyond what I could have imagined even just a year ago. I know I’m not the only one exploring, and it feels fun to be part of a community figuring out where we all go from here, together.\n\n## I’m the bottleneck\n\nIn just the last few weeks I’ve found two things:\n\n1. The primary constraint on me doing even more is my own limited capacity to manage large quantities of agent threads.\n2. I can delegate to agents (even ones powered by mid-tier models) that do a much better job than I do at prompting coding agent teams, managing endless agent threads, merging code, building, releasing, and so on.\n\nIf I want to build more and go faster, I have to let go of the minutiae. Agents handle the details better, don’t get tired, and don’t have other things to do IRL. At least for side projects, I get to be the CEO of my own eager team of bots: set direction, judge results, give feedback, and decide when we’re done.\n\nThat’s still a job. As powerful as that $200 Claude Code sub is, it does absolutely nothing unless I start talking to it.\n\nSo, I know I’m the bottleneck. I managed to accumulate several computers, a stable of side projects, and more agent threads than I could ever manage. Someone has to start work on the right machine (working on an iOS app only makes sense on my macOS boxes, not Linux), notice when a coding agent finishes a turn, read what it did, answer questions, decide what to do next, and so on. I’m probably still the best at doing that on one thread, maybe up to five. But that’s when things break down.\n\n## A bot for every project\n\nBut, there is another way. With Grok Bot (or any long-lived, project-based agent system you can delegate to) I hand off all of the above by default and only get brought in for truly critical decisions, where I define what is critical. Grok Bot understands what I mean and deals with the boring work of prompting agents in T3 via my gateway, so the right work is scheduled with the right coding agent on the right machine and seen through as far as it can possibly go without me.\n\n## The moving parts\n\n### The foundation\n\n- **[T3 Code](https://t3.codes/)** : Like Claude Code or the ChatGPT (Codex) desktop app, but it wraps almost any model, plus a background process with an MCP server that you can put on any machine to use your subs, plus a great mobile app, etc. Unless you just can’t ever leave the terminal, I’m not sure why anyone would use anything other than T3 right now.\n- **[T3 Fleet Gateway](https://github.com/boundsj/t3-fleet-gateway)** : My own MCP wrapper around what T3 offers natively. It makes it a bit easier to connect a hosted agent (like Grok Bot) to manage agents in T3. I wrote about[how it works](https://www.bounds.dev/posts/t3-fleet-gateway/) last month.\n\n### The project manager\n\n- **[Grok Bot](https://x.ai/bot)** : It’s many things. But I like to think of it as a[mid-tier model](https://x.ai/news/grok-4-7) that I would never code with, inside an OpenClaw-style outer control and management loop with routines (timed routines), skills, memories, and so on.\n\nWith just these ingredients, I have (IMO) the best system to wrangle coding agents: T3, my T3 Fleet Gateway routing system to schedule work on Codex and Claude Code agents on all of my computers, and Grok Bot to create infinite, per-project bots that manage all of it.\n\nT3 with my gateway is amazing, but Grok Bot is the secret ingredient. It lets me be the CEO. I can let go of managing every thread, delegate the work, and give feedback on the results.\n\n## Then the usage crisis\n\nBut then the usage crisis happened. Sometime last week, for whatever reason, Grok Bot usage started burning faster. That started me thinking about what I would want to change besides usage, and that line of thinking left me wondering: why not make my own Grok Bot? So I worked with Grok Bot to create a goal prompt to recreate itself. And [Botons](https://github.com/boundsj/botons) was born!\n\nThe great thing about Botons is that it takes the part of Grok Bot that changed how I build, the ability to delegate to a persistent team, and makes it mine. I can choose the models, connect the tools, and give each Boton a role. Instead of managing a pile of coding threads myself, I can talk to my team, set direction, and let them handle the details.\n\nJust like with Grok Bot, I can tell a Boton what I want to build and let it figure out where the work should run, brief the coding agents, check what they produce, and keep things moving. When it needs a decision only I can make, it comes back to me. I don’t have to sit there watching every thread.\n\nFor a sense of scale: between October 7, when my gateway went live, and November 12, it recorded 8,461 events from 376 jobs it started or tracked. Jobs started by Grok Bot account for 5,754 of those events. Add the jobs from the Minions projects, where Botons is being built, and the outer loop accounts for 87%.\n\n## The numbers, and how to read them\n\n| Started by | Events | Jobs | \n|---|---|---|\n| Grok Bot | 5,754 68.0% | 293 77.9% | \n| Minions/Botons | 1,614 19.1% | 61 16.2% | \n| Adopted/standing | 946 11.2% | 10 2.7% | \n| Other/test | 147 1.7% | 12 3.2% | \n| **Total** | **8,461** | **376** | \n\n- Events are gateway ledger records, not completed tasks, and leave out work done in T3 outside the gateway.\n- Events and jobs count toward the client that started the job, not necessarily the one that sent a later follow-up or answer.\n- Grok Bot is every bot on its shared connector. Minions/Botons is another gateway client, used in the Minions projects since November 5; Botons is the product being built there, not a separate client in the data.\n- The 376 jobs are 366 the gateway launched plus 10 existing threads it tracks. Outer-loop systems can drive those too, so 87% is the two named buckets, not an exhaustive measure.\n- Other/test is early pilots, QA and a few one-offs.\n- Snapshot from about 8:50 AM PT on November 12, 2026; the gateway went live at 2:06 PM PDT on October 7.\n\nI’m still figuring out where all of this goes. But for now, I’m having a lot of fun seeing how far my own bot control plane, with me as CEO, can take my ideas.", "url": "https://wpnews.pro/news/building-with-grok-bot-in-the-outer-loop", "canonical_source": "https://www.bounds.dev/posts/building-with-grok-bot/", "published_at": "2026-11-12 17:25:00+00:00", "updated_at": "2026-11-12 17:53:49.872316+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "agent-protocols", "developer-tools"], "entities": ["Grok Bot", "xAI", "T3 Code", "T3 Fleet Gateway", "Claude Code", "Codex", "OpenClaw", "boundsj"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/building-with-grok-bot-in-the-outer-loop", "markdown": "https://wpnews.pro/news/building-with-grok-bot-in-the-outer-loop.md", "text": "https://wpnews.pro/news/building-with-grok-bot-in-the-outer-loop.txt", "jsonld": "https://wpnews.pro/news/building-with-grok-bot-in-the-outer-loop.jsonld"}}