Most people who run multiple AI coding agents discover the same failure mode: the agents are individually fine, and the system is chaos. Two sessions edit the same repo and collide. A long-lived chat quietly degrades until it starts re-litigating decisions it made last week. A bot review gets skipped and nobody notices until the bug ships.

None of this is fixed by better prompts. It’s fixed by an operating model — and the core of one installs in about 90 minutes, on one repo. Do it on the repo where agents are already causing you pain.

The premise you have to accept first: one human merge authority. Agents never merge and never deploy. If that sounds too restrictive, you haven’t had the incident yet; I have, and rule one exists because four must-fix bugs once shipped in a single “implement, admin-merge, deploy” pass that no human gated.

0:00–0:20 — Install the fleet rule

Write down the behavioral invariants every agent session must follow, and put them where your editor injects them into every new chat (for Cursor, an always-on rule in .cursor/rules/). Two things matter more than the wording:

  1. Name the review bots you actually run — “wait for review” is unenforceable; “wait for Bugbot and CodeQL” is a checklist.
  2. Delete anything that contradicts your real workflow. A rule agents watch you break trains them to ignore all rules.

From this point, every new chat starts with the fleet’s ground rules instead of relearning them mid-incident.

0:20–0:50 — Stand up the merge gate

Create a merge checklist and put it somewhere you will actually see it: repo wiki, CONTRIBUTING.md, or printed next to your keyboard. The essential line is the expected bots contract: which automated reviewers must report before anything merges. This turns “the checks look fine” into a definition of done — and makes a silently skipped bot visible, which is the failure that bites hardest.

Then add a PR template so every agent PR arrives with acceptance criteria, out-of-scope notes, and a statement of which other in-flight work it might overlap. Overlap notes are what save you when two sessions touch one repo.

0:50–1:10 — Set your health thresholds

Agent sessions degrade with context, and the degradation is measurable before it’s visible. Adopt three states now, tune later:

  • Under ~60% context — work normally.
  • 60–80%, or a very long thread — finish the current unit of work, then write a handoff brief.
  • Over ~80%, or visible confusion — stop new work and hand off now.

Write the numbers where you’ll see them. A manual glance at the context meter at the start of each work block is enough to begin; I eventually built a small CLI so the measurement happens automatically.

1:10–1:30 — Write your first handoff brief

Pick your longest-running chat — the one you’ve been nursing because “it knows the codebase.” Write a brief that captures: the goal, the state of the work, and above all the decisions locked (so the next session doesn’t reopen them). Then kill the chat and start a fresh one from the brief.

This step is the point of the exercise. Rolloff only becomes normal operations after you’ve done it once and watched the fresh chat pick up the work without re-litigating anything.

What you deliberately skipped

Leave these for the first week, once the gate and the health line are habits: a critical-path list (which repos touch money or authz, and the extra test gates that hang off it), a knowledge system (distill playbooks after your first few handoffs, when you can see what keeps getting re-explained), and MCP credential hygiene (which matters before you grant agents write access to real systems — not before).

You’re running the model when: agent PRs arrive with the template filled, no PR merges with an untriaged or silently skipped bot, and you’ve rolled off at least two chats without losing work. Scale agent count after that — not before.

This is the quickstart chapter of Running a Fleet of Coding Agents — a book plus the actual templates (fleet rule, merge checklist, PR template, handoff brief) so the 90 minutes is copy-and-edit, not blank-page. The LAUNCH49 link takes $30 off while the launch window is open.