GlossaryGlosarium
What is an agent fleet?What is an agent fleet?
An agent fleet is a group of AI agents that share one rulebook, one human approval seat, and one append-only log. Each agent covers a different class of work — event operations, member engagement, moderation — and none of them can act on its own.An agent fleet is a group of AI agents that share one rulebook, one human approval seat, and one append-only log. Each agent covers a different class of work — event operations, member engagement, moderation — and none of them can act on its own.
Glossary / What is an agent fleet?
An agent fleet is several AI agents working inside one shared set of rules. The shared part is the point. A single agent has to be trusted wholesale: either you let it write to your community or you do not. A fleet splits the work into named classes — promoting an event, chasing an RSVP, assembling a recap, moderating a thread — and each class carries its own permission.
Three things are shared across every agent in the fleet. First, the rulebook: one pack that declares which action classes exist, which tools are available, and what each class is allowed to do. Second, the human seat: one named person who decides the proposals. Third, the log: every proposal, verdict, refusal and grant is appended to a single replayable history.
The shared rulebook is what removes the failure mode that makes community AI tools feel like spam. A single bot that can post, reply, and DM will eventually do all three at once, in the wrong place, at the wrong hour. A fleet keeps the blast radius small, because a class that is not declared does not exist and a class set to draft never runs on its own.
The agents are not interchangeable, and they are not personalities. In the shipped event-operations pack they are role-shaped — event operations, engagement, moderation — and each role holds a different toolset. The moderator can read and flag; it holds no delete, no ban, and no mute, because those tools are absent from its toolset rather than configured off. Absence is the design: a reviewer can read the pack and see what the fleet cannot do.
Related
| Term | |
|---|---|
| What is the human seat?The human seat is the one named person allowed to decide a proposal, issue or revoke a grant, and release a freeze. Every member-facing message the fleet produces passes through it first. It is a role, not a password, and the control plane ships with exactly one. | Read → |
| What is per-action autonomy?Per-action autonomy means every class of work gets its own rung on a three-step ladder — draft, opt-in-auto, or never — instead of one switch for the whole agent. An unset class behaves as never, never is a floor, and the narrower of the pack and the organisation always wins. | Read → |
| What is a replayable audit trail?It is an append-only log where every proposal, verdict, refusal and grant is written once and never edited. Because state only ever changes by folding events in, replaying the log through the same reducer reproduces the state exactly — and the check is run, not assumed. | Read → |
By Filbert Then · Last updated · Suggest a correction
See the fleet runLihat armadanya berjalan
Run the seeded cycle, decide the proposals, then read the audit trail.Jalankan siklus contoh, putuskan usulannya, lalu periksa audit trail-nya.