Run a multi-agent build

Step 1 — Write the shared truth

Multi-agent work has one secret, and it isn’t the agents: windows coordinate through the source of truth, not through each other. Nobody relays messages. Everyone reads and writes the same plan. In the project folder, have your first window draft it:

Create (or update) plan.md as a coordination file for parallel work: THE GOAL, one paragraph. THE WORK QUEUE — three tasks: [a build task — a feature or fix; a content task — rewrite of the words; an audit]. OWNERSHIP — which files each task may touch; no file gets two writers. DECISIONS — an empty log we’ll append to. Keep it under one page.

The plan file is the meeting room. Think of ten designers on one launch: the logo person isn’t in the copy person’s work — but everyone can see the wall.

Step 2 — Cast the windows: one job each

Open three windows, each scoped to a job, not a topic:

• The BUILDER — Claude Code, in the folder: owns the build task and the repo.
• The WORDSMITH — the chat window: owns the words, and receives only the copy brief and the current text, never the whole repo.
• The INSPECTOR — a separate session (a different tool if you have one — a second brand of skepticism), fresh, and read-only by instruction.

Why split at all: one window doing everything fills with everything — and rots. Three scoped windows each stay sharp, because each carries only its job. And note what you’ve just built: three agents — window, instructions, reach — so ask that entry’s four questions of each: what’s in its window, what can it touch, where’s its truth, who reviews.

Step 3 — Run the builder and the wordsmith in parallel

Kick off the builder:

Read plan.md. Take the build task. Touch only your files per OWNERSHIP. Commit as you go; when done, append what you did and any decisions to plan.md’s DECISIONS log.

While it works, run the wordsmith in the chat — paste the copy brief and current text, get the rewrite, and drop the result into the folder (or hand it to the builder via a plan.md note). This is the rhythm the whole lab exists to teach: set one off, work the other, come back. You are not waiting on AI; AI is waiting on you.

Between tasks, you do the orchestrator’s only jobs: update the queue, log decisions, merge results.

Step 4 — The inspector: independent and read-only

When the builder settles, the fresh window gets the whole repo:

You are a read-only inspector. Read the repository and plan.md. Do not edit anything. Report: does the build match the plan? What would break for a careless user? What did the builder assume that plan.md doesn’t actually say? Rank findings by severity.

This is the fresh-window move from Red-team critical work at system scale — the builder’s window is invested in its own work; the inspector’s isn’t. Read-only is the discipline that keeps it honest: an inspector who can edit becomes a second builder, and now nobody is checking. You triage the findings into the queue — not everything deserves fixing — the builder fixes, the inspector verifies the diff.

Step 5 — Converge, ship, and leave breadcrumbs

Merge the words, run the fix batch, deploy. Then the close that makes tomorrow cheap:

Append to plan.md: DECISIONS made today, open items, and a three-line journal entry — what was done, why, and where things stand.

Tomorrow’s windows start warm because the truth lives in the folder. The thread never lived in your head — which is why you never lost it.

Now name what just happened: scoped windows, per-role loading, role instructions as steering, one source of truth, agents with jobs and reach, and independent verification. That’s the entire reference library, running at once, conducted by you. The fancy orchestration frameworks automate the queue; the shape is the one you just ran by hand.

---

Check your understanding: how do the three windows coordinate without ever talking to each other — and why is the inspector forbidden to edit?

Use this process for other complex work. Researcher–writer–checker on the big report. Builder–designer–copywriter on the launch. Analyst–drafter–red-team on the deal. Any work too big for one window was always going to be a team — now you know what the team is made of.