Seven practices, each derived from something that runs.
Seven practices for running a fleet of AI coding agents, each derived from a working artifact rather than written as advice. Every page names the file it comes from and, where it counts something, the command that recounts it.
- The operating manualHow a layered set of CLAUDE.md and AGENTS.md files gives every agent the rules for the directory it is standing in.
- WorktreesWhy every branch gets its own working tree, where they are allowed to live, and the two-line helper that enforces it.
- The fleetHow a 132-line shell script keeps parallel agent sessions findable instead of scattered across terminal tabs.
- The session catalogA local Swift daemon that indexes agent transcripts into a read-only catalog and serves six search tools back to the agents that wrote them.
- VerificationAn adversarial review layer that attacks findings on four fixed lenses and fails closed when it cannot reach a verdict.
- Eval-gated skillsA loop that edits an agent skill, measures the change against a held-out eval set, and refuses to promote it without a statistical margin and an independent verdict.
- The attention ledgerOne file defines every metric, three reporters read it, and the hardest work was removing the automation from its own numbers.