A measured doctrine for orchestrating AI agents, and knowing when not to.
Most "multi-agent" advice assumes more agents is better. It usually isn't. Fleet Mode is the opposite stance, run as a live Claude Code skill: adding agents has a negative average payoff on most tasks, so you fan out only where it demonstrably earns it, and you keep the writes single-threaded and gated.
It's the operating mode behind bow (an autonomous all-Claude chief-of-staff agent) and the builds it ships.
🧩 One layer of a five-repo cost-governance stack for operating AI agents cost-efficiently; bow is the flagship that runs every layer in production.
- Read-heavy work fans out. Research, codebase/PR review, multi-file audits: parallel subagents in clean contexts, each returning a condensed summary. This is the only thing fan-out is for.
- Writes stay single-threaded. One agent makes the edit. Never fan out to edit in parallel.
- Deterministic checks first, then an independent refute-first review. Tests/types/lint/build must pass with no model in the loop; then a separate clean-context reviewer tries to refute the work. No agent grades its own work. Fail closed.
- Every added agent earns its fan-out. Default to a single strong agent. More agents add intelligence (extra perspectives, verification), not parallel actions.
Plus the two non-negotiables: human-gate anything irreversible/MAJOR (push, deploy, send, real-money trade, mass-delete, new spend), and log an honest receipt of every kept/killed decision with the real number.
- Classify stakes. trivial: light check, ship. non-trivial: full gate. irreversible: full gate plus human approval first.
- Decide fan-out (bias-to-NO). Single strong agent by default; fan out only for read-heavy, parallelizable work that exceeds one context window.
- Write single-threaded.
- QC gate. Deterministic checks, then an independent reviewer that tries to refute. Escalate high-stakes output to a different-model judge.
- Human-gate MAJOR items.
- Log a receipt.
Full operational spec: SKILL.md.
The skill references the superpowers:* skill family (for dispatching parallel agents and
requesting independent review). If superpowers skills are unavailable, substitute: fan-out
decisions manually, run tests before any model review, get a colleague to review independently.
The receipt script has zero dependencies (Python stdlib only).
git clone https://github.com/Jott2121/fleet-mode ~/.claude/skills/fleet-modeClaude Code auto-discovers it. Invoke it (or let it auto-apply) on any non-trivial change. The receipt tool stands alone too:
python3 scripts/append_receipt.py \
--task "add retry to uploader" --tried "exponential backoff" --verdict kept \
--why "cut timeout errors to zero in a 200-run soak" --metric errors --value 0Appends an append-only row to KILLLOG.md: the honest ledger of what you kept, killed, and why.
See KILLLOG.md for a sample of the format.
Evaluate without prejudice; adopt only on measured proof. Receipts over hype.
The multi-agent discourse is full of impressive-sounding fan-out diagrams that, measured, lose to a single careful agent. Fleet Mode is the discipline that keeps the wins (parallel reading, independent review, hard gates) and drops the cargo-culting (parallel writers, self-grading, unbounded agent swarms). It scales to the task, not to the hype.
A doctrine about hard gates runs its own:
- Coverage-gated tests — the receipt-writer script is tested on Python 3.11–3.12 and the build fails below the coverage floor (currently 92% covered).
- CodeQL —
security-extendedstatic analysis on every push, PR, and weekly; findings surface in the Security tab. - Pinned supply chain — GitHub Actions pinned to commit SHAs, kept current by Dependabot.
- Branch protection —
mainrequires CI + CodeQL to pass before a merge. - Disclosure policy — see SECURITY.md; private reporting is enabled.
MIT, copyright 2026 Jeff Otterson. See LICENSE.