The agent team and the brain

The team

I do not run one big agent. I run one orchestrator and a few workers. The orchestrator plans, hands out the work, and checks every result. It does not write product code. Each worker does one job, with only the tools that job needs.

• Researcher: answers one question with sources. Never edits.
• Builder: makes the failing tests pass, inside the files the brief names.
• Tester: writes the failing tests first. Later, it checks the build on its own.
• Quality checker: tries to break the builder's claim. Says PASS or FAIL. Never fixes.
• Ops: runs the git, GitHub and ticket steps the brief lists. Merges only when the owner said so.

We also keep a small visual checker. It looks at screenshots and sends back words, so images stay out of the orchestrator's memory.

You can set up this team in one paste. The R.B.T.Q. agent team prompt writes the files. The Orchestrator Course explains the whole method.

How the team uses the brain

Workers start fresh. They do not see what the orchestrator read. So the brain flows one way: down from the orchestrator, then back up as plain facts.

1. The orchestrator reads the brain once, at the start of the task.
2. It puts the brain facts each worker needs into that worker's brief. That means the brain commit, the page IDs it read, and short quotes of the facts that matter.
3. Workers never write to the brain. If a worker finds a fact has changed, it says so in its report.
4. At the end, the orchestrator writes one update for the whole task.

Why one update per task

Five workers writing five updates is noise. The librarian would get five half stories. One update from the orchestrator tells the whole story: what changed, why, and how it was checked.

Set it up

1. Set up the agent team with the R.B.T.Q. prompt, or use your own team.
2. Paste the prompt below into Claude Code in your repo. It adds the brain rules to the orchestrator.
3. Start a new session, so the rules load.
4. Give the orchestrator one small task that touches a company fact.
5. Read one worker brief. Check that it carries the brain commit and the facts.
6. At the end, check that only one update was written, by the orchestrator.

The prompt: add brain rules to the team

Add company brain rules to the orchestrator in this repo.

1. Read this repo's CLAUDE.md. Keep everything already in it.
2. Add this block to the end of CLAUDE.md:

   ## Brain rules for the agent team
   - At the start of a task, read the company brain once, using the rules link in the
     brain section above. Write down the brain commit and the page IDs you read.
   - Every worker brief carries the brain facts that worker needs: the brain commit,
     the page IDs, and short quotes of the facts that matter.
   - Workers never write to the brain. If a worker finds a changed fact, it reports it back.
   - At the end of the task, only the orchestrator writes the brain update. One task,
     one update. If no fact changed, say "No material company fact changed."

3. Read each agent file in .claude/agents/. In each one, add this line under its rules:
   "Never write to the company brain. Report changed facts back to the orchestrator."
4. Show me every change. Commit on a new branch and open a pull request. Do not merge it.

You are done when

• Worker briefs carry brain facts.
• Workers report facts back and never write to the brain.
• Each task ends with one update, or with "No material company fact changed."

Next module

Module 7 is the daily habit that keeps the brain true: ten minutes a day for the owner.