How work is divided
The lead agent (the chief of staff) gets each request. It selects one of three paths by the size of the job.
Inline
A small job (a fix, a cleanup, a config change) stays with the lead agent. It runs the gates, pushes, watches CI and tells the owner after.
One worker session
A job with more than one step in one repo goes to one worker. The lead agent writes a brief and runs
fleet spawn <name> --project <repo> --workdir <worktree> --model <model>. The brief is the whole
context of the worker. It gives the goal, the limits, the done condition and how to report. The
lead agent sets a watch on the worker's state in the same step. A timer for a guessed time is not
a watch.
Build pipeline
| # | Stage | Who | Output |
|---|---|---|---|
| 1 | Go or no-go | lead agent and owner | One paragraph: why, scope, alternatives |
| 2 | Spec | one worker and research subagents | docs/spec.md with open questions |
| 3 | Rulings | lead agent (technical), owner (product) | Rulings in the spec, which is then frozen |
| 4 | Breakdown | a new session, not the spec author | Seams, waves and one issue per ticket |
| 5 | Wave 0 | one session | A walking skeleton and the contracts as code, deployed |
| 6 | Swarm | one worker per ticket | Merged tickets, one at a time |
| 7 | Review | an agent that did not write the code | Pass, or findings back to the ticket |
| 8 | Integrate | lead agent | Merge in dependency order, deploy, report |
- Seams are contracts written as code: an OpenAPI file, a schema, an interface. Wave 0 builds them. A worker that finds a contract incorrect stops and reports blocked.
- Waves show dependencies, not phases. A ticket starts when its own dependencies are on main. Two tickets that change the same file are one ticket, or the file is a seam for wave 0.
- A ticket owns paths. A worker edits only those paths.
- Tests come first. A ticket is done when its tests and the gates pass, not when the worker says so.
- Size the swarm to the usage budget, then to the memory of the box.
docs/fleet/breakdown.md is a worked breakdown with seam tables.
Model choice
| Work | Model |
|---|---|
| Lead agent, spec, breakdown, review of a wave | Opus |
| Wave 0 and a ticket that changes a shared contract | Opus |
| A ticket with fixed contracts, owned paths and acceptance tests | Sonnet |
| Small follow-ups, tests only, text | Sonnet |
| Architecture of a new large project, a problem Opus failed on twice | the most capable model |
Go up one step (Sonnet, then Opus) when a ticket stalls twice on the same problem, asks to change a contract, or review finds a real defect. Record the change on the ticket.
Review
Another agent reviews each ticket. It checks the diff against the acceptance criteria and the spec. It checks that the tests can fail and that no file is outside the owned paths. It sends findings back to the ticket's worker. It does not write cosmetic findings.
Report
Every report gives evidence: the command and its output, or a link. A claim that is an inference says so. A closing comment has three lists: proven by fakes, proven live, not proven.