This week one chat message cost me most of a usage window.
I run a few dozen Claude Code sessions on a Mac mini and talk to them from my phone. Each one has a job: a lander that merges branches, builders, a reviewer, a couple of content pipelines. When the rate limit reset, I sent the same line to all of them: "usage reset, continue".
About forty agents woke up at once. Seven minutes later, 83% of the window was gone.
Nothing was wrong with any single agent. The system had no rules for running many of them. Here are the five rules that came out of that and a few failures like it. They are plain, and each one exists because of a specific failure.
1. Lanes
An agent that notices unfinished work will try to finish it. With many agents sharing a repo and a notes folder, that gives you double builds and agents "fixing" each other.
The rule: each agent sticks to the task it was briefed on. If it spots something for another project, it writes one line in an ideas file and carries on. One manager session is the only thing that routes work between agents.
The sign it works is the boring report: "both features already exist, nothing committed, stopping."
2. Write to the shared notes always, read them by grep
My first rule was "read the shared notes at the start of every session". That burned context in every session, every day, for things most agents did not need.
Now: writing is mandatory (handoffs, lessons, decisions), and nobody opens those files whole. Unclear task? Grep the handoff file for the topic and read that section. Cap the files and give one agent a weekly trim.
3. Builders never push to main
Letting every agent run the full test suite and push fails twice: the machine melts under parallel end-to-end runs, and main breaks anyway because each agent tested against a different main.
Builders work in their own worktree, run only the tests they touched, push a branch and send the sha to one lander. The lander batches two or three branches, runs the gate once, and pushes. A pre-push hook refuses main without a green gate. Heavy jobs (browsers, e2e) go through a lock directory with an owner file, one agent at a time.
4. Usage is the budget
- Wake agents in waves after a reset: lander first, then two or three builders, then the rest.
- Track used% against elapsed% of the window.
- At 85 to 90%, everyone pushes what they have, writes a state file, and stops. Only the lander keeps going, so finished work is not stranded.
- Long reading (logs, PDFs, research) goes to a cheaper model that returns a short summary.
- A screenshot stays in context for every later call. One visual check at the end of a task, not one per tweak.
5. Reports say what was not checked
Every turn ends in one of three states: urgent, question, or done. No progress pings.
The most useful line in a report is the one about what was not verified: "pushed, unit tests pass, nobody has looked at it in a browser". An agent's mistakes arrive in the same confident voice as its results, so the report has to carry that itself.
One more that surprised me: "I'll check back later" does nothing. A session that ends its turn is gone until something messages it. The only real follow-up is a scheduled message into its own chat.
If you want the templates
I put the full rule set and seven copy-paste templates (global CLAUDE.md, handoff, lessons, agent brief, landing protocol, pacing) into a short paid PDF: The Agent Fleet Playbook, US$12. The five rules above are the core of it, so take those and go if that is all you need.
Disclosure: this post was drafted by one of the agents in that setup, from its own rules and logs.
What broke first when you went from one agent to several?

