agent-fleet
Rules for multi-agent work: model routing per phase, fan-out sizing, closed output schemas, and verifying what agents return.
/plugin install agent-fleet@4nrry
The skill lives in skills/agent-fleet/ and the plugin ships an Agent Plugins
plugin.json at its root, so Copilot, Codex and Cursor can install it; other
harnesses get the skill folder through link_skills.py. The hooks and the
workflow are Claude Code only. Paths below are relative to the skill folder,
except hooks/ and workflows/ at the plugin root.
| Component | Path | Does |
|---|---|---|
| Skill | SKILL.md | The rules, in tiers rather than model names |
| Harness map | references/harnesses.md | Tiers, model setting and read-only children per harness |
| CLI workers | references/cli-workers.md, scripts/fleet_worker.py | Search and extract on another CLI through crew's ask --web --schema, read-only |
| Verifier | scripts/verify_citations.py | Fetches each cited URL and string-checks the quote |
| Workflow | workflows/deep-research.js | Claude Code's built-in deep research with per-phase model: overrides |
| Injection hook | hooks/inject.sh | UserPromptSubmit: loads the skill when the prompt contains deep-research or ultracode |
| Routing hook | hooks/check-routing.sh | PreToolUse on Workflow: denies an unrouted script once, and the bare built-in name always |
Why hooks
The skill's description fired on 2 of 30 should-trigger runs, and no rewrite helped, so a substring hook loads it instead. The routing rule was loaded during a ~97-agent run that still inherited the top model, so a hook checks for it at dispatch.
Evidence
Records and claims live in bench/plugins/agent-fleet/:
the trigger eval, the hook-grep batch, the breadth-tier batch, and the
fleet-worker smoke notes. The rules in the skill come from earlier paired runs
that predate the bench protocol; read those as design notes.
workflows/deep-research.js is the 2.1.306 built-in plus model: overrides;
its header says how to rebase it when the built-in changes.