One of three roles in the catalogue. Agent mechanisms govern the fleet and the substrate that produces work: how briefs are composed, how agents are dispatched and isolated, how their commits are gated and merged, how shared host compute is rationed, and how the fleet's lifecycle is observed. (What the fleet reasons through is the models-bridge; what it produces is the product.)

1. Context & dispatch substrate — what an agent knows and how it is launched: brief-linting, the rule index, dynamic context injection, role-typed dispatch.
2. Gates & merge-train — the path-to-production staircase (pre-commit → sentinel → merge-train → deploy): defense in depth, each stair a backstop for the last.
3. Mediators & resource locks — host-level wrappers that ration shared compute across concurrent worktrees, each enforced so the raw call is impossible.
4. Lifecycle & observability — live signal surfaces over the fleet: the agent-registry, the typed event bus, heartbeats, tombstones, the cron-alerts gate.
5. Governance-doc mechanisms — documentation treated as enforced infrastructure, including the CLAUDE.md rule index — the meta-mechanism that records and delivers every other mechanism.

Soft aims, hard holds. Of the agent mechanisms, only dynamic context injection is purely soft (it injects context but blocks nothing); two are soft·hard (a document kept honest by a lint); the rest are hard deterministic gates. Probabilistic guidance is used to aim agents at the right work; deterministic machinery holds the line. See the umbrella README's move and form axes.

The census, forms, soft/hard axis, relationships, and entry template are documented once in the umbrella README.