Your bot is the manager, not the coder

1. 011:35

   The bot manages, the agents code

   A chat bot is good at deciding what to build and checking what came back. Writing the code is the job of a coding agent like Claude Code, Codex or Cursor.

   More
2. 021:38

   First mile, last mile

   The first mile is working out what to build from everything your bot can see. The last mile is chasing the work, checking the proof and telling you what's done.

   More
3. 032:23

   Not tied to one bot

   Rob uses Grok Bot, but says your own agents, Claude Projects or Cursor Projects work the same way. The pattern is the point.

   More
4. 043:36

   Split the jobs, split the bill

   Coding inside the bot burns its shared usage pool fast. Rob hands the coding to his Claude Code subscription and Cursor cloud agents instead.

   More
5. 053:54

   Every worker gets its own computer

   Each coding agent runs on its own machine with its own context and its own off switch, so one bad run can't wreck the others.

   More
6. 065:08

   A loop that runs every day

   Lauren Tan's five steps: watch a Slack channel, file a ticket, an agent reproduces the bug, fixes it and fuzzes it, then pings her and merges an hour later.

   More
7. 076:15

   Not every bug gets the loop

   Sort each change by blast radius. Green merges itself, yellow gets an AI review, red needs a person. Payments are always red.

   More
8. 086:36

   Head chef, not factory owner

   Lauren prefers a Michelin kitchen to a factory. You don't cook every dish any more, but your name is still on every plate.

   More
9. 097:08

   Trust before more agents

   How can you manage 50 agents if you can't trust one? The teams that ship this way started with checks, not with more bots.

   More
10. 107:32

    Move the fix up the stack

    Every time you correct an agent, fix it higher up: in the code first, then a lint rule, then a rules file. Reading the code yourself comes last.

    More
11. 119:25

    One chat, narrow bots

    Rob talks to one main bot, Chief. Behind it sit narrow bots, each with its own small set of permissions. He tries hard not to build 50.

    More
12. 1211:51

    Score the backlog

    His bot saves every idea to a Notion backlog, then a routine ranks it by RICE (reach, impact, confidence, effort) and marks how much could break.

    More
13. 1313:17

    All your bots share one computer

    Log in a coding tool with an API key on the bot's computer and every bot you have can use that key. Keep the rules in the environment, not in the bot.

    More
14. 1414:29

    Start with one rule

    Don't build 50 agents. Paste one rule into your bot: hand coding to an agent with a test, check the proof, open a pull request, report back.

    More

Manager bot (orchestrator)

The one assistant you talk to. It plans, hands work to other agents, checks what comes back and tells you. It doesn't write the code itself.

Coding agent

An AI that edits code and runs commands in a project: Claude Code, Codex, Cursor. On this page it's the worker.

Cloud agent

A coding agent that runs on a computer in the cloud instead of your laptop, so it keeps going when your lid is shut.

Headless

Running an agent with one instruction and no chat window. It does the job and stops. This is how a bot starts Claude Code on your Mac.

First mile

Working out what to build at all: reading Slack, email, tickets and complaints, and turning them into a short list.

Last mile

Following the work through: chasing it, checking the proof, opening the pull request and reporting back to you.

Intent

What you and your team actually want built. Rob's point: a big part of any software factory is collecting and managing it.

Routine

A job your bot runs on a schedule (every morning) or when something happens (a new Slack message). Grok Bot's word; others call it an automation.

Backlog

The list of everything that could be built or fixed next, usually in Notion, Linear or Jira.

Ticket

One item in a tracker like Linear or Jira: a bug or a feature, with a title, a description and a status.

Pull request (PR)

A proposed change to the code, waiting to be checked and accepted. Opening one is safe; merging it changes the real thing.

Merge / auto-merge

Accepting a pull request into the main code. Auto-merge means it happens on a rule, with nobody pressing the button.

Acceptance test

A check that proves the job is done, written before the work starts. "The total can't go below zero" is one.

Proof

What the agent hands back to show the work is right: passing tests, screenshots, a short video of it working.

Reproduce (repro)

Making a reported bug happen again on purpose. If you can't make it happen, you can't prove you fixed it.

Fuzz

Throwing odd and random input at something to see what breaks: tapping minus ten times, pasting emoji, double-clicking Pay.

Swarm

Several agents working on the same thing at once. In Lauren's loop, a small swarm tries to break each fix.

Playwright

A free tool that drives a real browser from code: it clicks, types and checks the page like a person would.

Blast radius

How much could break if this change is wrong. A typo fix is small. Anything touching payments or logins is big.

RICE

A way to rank work: Reach (how many people) times Impact times Confidence, divided by Effort. Small jobs that help many people rise.

Static analysis

Checks that read your code without running it: lint rules, type checks. Cheap, fast and the same every time.

Lint rule

An automatic check for one specific mistake, like leftover console.log lines. It fails the build so nobody has to spot it.

CI (continuous integration)

A service that runs your tests and checks on every pull request, before anything is merged.

Rules file

A plain text file of instructions an agent reads before it works: AGENTS.md, CLAUDE.md. The video calls this layer prose.

Skill

A saved, named set of instructions an agent can load for one kind of job, like reviewing code or writing a test.

Revert

Undoing a change after it was merged, usually because it broke something.

Vault

Where a bot keeps your passwords and keys so it can log in for you. Rob can't see inside Grok Bot's, and that bothers him.

Scoped credentials

Keys that only open what one job needs, and read-only where possible. A key that can read logs but can't delete anything.

Sandbox

A walled-off computer or container where an agent can work without reaching the rest of your machine or accounts.

VPS

A rented always-on computer in a data centre. Rob's bot can start agents there as well as on his Mac.

Container

A sealed box on your computer with its own files and tools, so an agent inside can't touch the rest.

MCP

Model Context Protocol: a standard plug that lets one AI app call tools in another. Theo drives his machines through one.

1. 1 · Manager botFirst mileReads every channel, works out what to build next, writes the job up and picks a worker.
2. →
   2 · Coding agentsThe diffClaude Code, Codex or Cursor writes the code on its own computer and brings back proof.
3. →
   3 · Manager botLast mileChecks the proof, opens the pull request, chases what's stuck and tells you what's done.

Lauren Tan, who works on Grok Bot, puts the middle step in her own words: since each cloud agent has its own computer, “it frees up your bot to be more of a manager rather than write the code itself.” Her post. In another, her last mile also covers handing the build to the coding agents; Rob draws the build as its own middle step.

The five steps of the loop are Lauren's, word for word in each station. Rob adds the triage colour and the review levels. Splitting them into seven stations and the example bug are ours. Anthropic uses low, medium and high, not colours: low merges once automated review and security checks are green, medium needs one person, high needs two. Their post.

1. 01

   Pick your manager

   Whatever chat assistant you already use that can start other agents or run commands: Claude Projects, Cursor, Grok Bot, or a Cloudroom or bb (getbb.app) thread. For a first try it can even be a second terminal window.

2. 02

   Paste the first rule

   Rob's one rule, word for word, goes into the manager's instructions. It stops the manager from writing code and makes it hand off, check and report instead.

   The first rule
3. 03

   Give it one worker

   Claude Code or Codex, run headless in your project folder. Try it by hand once so you know what the manager will be doing. The worker works on a new branch and opens a pull request; it never merges.

   Give it one worker

   ⧉Copy

   # Claude Code, one job, no chat window
   claude -p "Fix: the share button does nothing on Safari. Done when a Playwright test opens /share in WebKit, clicks Share and finds the page URL on the clipboard. Work on a new branch, run the tests, open a pull request, and reply with the PR link and the test output."
   
   # Codex, same job
   codex exec "Fix: the share button does nothing on Safari. Done when a Playwright test opens /share in WebKit, clicks Share and finds the page URL on the clipboard. Work on a new branch, run the tests, open a pull request, and reply with the PR link and the test output."

4. 04

   Run one small, green job

   Pick a fix nobody would lose sleep over. Write the acceptance test as one sentence. Ask your manager to hand it off, check the proof and report back with the pull request. You merge it yourself.

5. 05

   Every correction moves up

   When you have to fix the same thing twice, don't just comment. Ask the agent to turn it into a lint rule or a code change. This one stops stray console.log lines in a JavaScript project.

   Every correction moves up

   ⧉Copy

   // eslint.config.js: add to the rules of your existing config
   rules: {
     "no-console": "error",
   },

6. 06

   Lock the doors before you add a bot

   Give the worker its own keys, read-only where possible. Keep payment and admin keys off any computer your bots share.

   The door rules
7. 07

   Add a channel, then a colour

   After a week of clean pull requests, let the manager watch one channel (a Slack channel or a Linear view) and add the triage starter. Only then let green jobs merge on a rule.

   Triage starter

• pstack (Lauren / poteto) The skills her loop runs on, the Michelin kitchen and her full correction ladder with her own prompt.The skills behind steps 3 to 5
• In loop, near loop, out loop IndyDevDan's ladder: you prompt, you approve, you observe. Rob's green tier is near the loop.How far you step back
• Warp: software factories Crawl, walk, run, from you doing the work to agents doing it.Same idea, factory words
• Steve Sewell's factory A factory that fixes bugs on a timer.The bug loop, another way
• Ondrej's mini software factory Loops that check their own work and run around the clock, at small scale.A factory you can build alone
• Ras Mic's SDLC factory His software factory from plan to ship.Another person's whole line
• Compound engineering The plan, work, review loop. Alan Wells' three-person factory runs a fork of it.The worker's playbook
• Theo on token-maxing Ten threads at once, all watched by one person. Now he drives them from Grok Bot.In the loop, at full speed

A fair pushback, from a post Rob links: is this just cloud agents hooked to a few webhooks? Mostly, yes. The plumbing is simple. What isn't is the checks that let you stop reading every change, and Jellyfish's numbers say almost nobody has those yet.

• “Grok Bot + Claude Code, Why is Everyone Switching to This?” by Rob Shocks (Switch Dimension, posted Oct 31, 2026). The manager-not-coder split, the three reasons, the traffic light, his bots, the backlog, RICE with blast radius, the door rules and the first rule are his, in our words. Slide labels, his bot list, the backlog board and its review labels are read from the screen; where captions garbled a name, the screen won.
• Lauren Tan's loop, first and last mile, manager line, inverse pyramid and routines are from her posts on X and LinkedIn. She works on Grok Bot; Rob notes her job is to promote these tools. The Grok Bot guide credits her with 2,000+ pull requests in a month (x.ai).
• Disclosures: the video is sponsored by Neon (left out here). Rob sells the Switch Dimension course and community, plugged near the end, and is on a combined Grok and Cursor plan. Theo makes T3 Code; Rob says he isn't affiliated with Grok Bot.
• Checked on Oct 31, 2026: Anthropic's figures and quote, Docker's central idea, the Grok Bot shared-computer list, Cursor's shared pool, and the posts by Lauren, Theo, Alan Wells and hunvreus. Rob's green, yellow and red are his own framing; Anthropic's tiers are low, medium and high. Not checked: what Rob's AI review levels run, and his own results.
• The seven-station map, the example bug, the console.log fix at each layer, the basic examples, the try-it steps, the fill-ins, both starters, the sibling list and the takes are FakeNerd's. Not sponsored by anyone on this page.