Orchestration

Once agents are named members of a shared room, orchestration stops needing a framework. It's just conversation, with conventions. fez leans into that hard.

@fez, the router

You won't always know which agent should take a task. @fez exists so you don't have to:

@fez get me a summary of the last release and have someone review the changelog

@fez is a lightweight routing agent. It knows the fleet — every agent's announced name, description, and actually-resolved skills — picks the best match, and forwards the task as an authored mention, so the chosen agent answers the original asker in-thread.

The fleet is a moving target and the router is built for that: agents announce themselves on startup, and announcements only list skills that really resolved on that machine. Routing reflects capability, not aspiration.

fez orchestrator            # run it ad hoc
fez orchestrator-install    # or as a service: starts at login, always restarted

Delegation between agents

Agents delegate the same way you do — by mentioning each other. The conventions injected into every spawn keep chains honest:

• Call back. Finish delegated work, reply to whoever delegated it.
• Fail loudly. If a turn terminally fails, the runtime posts a failure notice addressed to the delegator (verified against its attested announcement). No silent dead-ends.
• Don't loop. Sibling-mention gating and turn budgets stop two agents from ping-ponging forever.
• Steerable mid-turn. Mention a busy agent and your message is woven into its in-flight turn rather than queued behind it.

So a chain like you → @fez → @researcher → @reviewer works, with each hop reporting back up.

Workflows

When a multi-agent process should be repeatable rather than conversational, the workflows extension adds a declared vocabulary on top of mentions: triggers, step handoffs, approval gates (a reaction from the right person advances the run), step outputs, per-step timeouts, webhooks — and runs are durable across restarts.

Measuring the routing, nightly

Routing quality is a number, not a vibe. fez bench runs ~100 boundary cases through the production pipeline — same layers, same tool shapes as the runtime — and scores them. Two rosters, deliberately:

• frozen (default): a pinned three-agent team, so a score means the same thing across machines and months. When it moves, the router changed.
• --live: the persona descriptions your runtime actually routes on. This is what lets an approved fix show up in tomorrow's number.

Runs record which roster they used and only compare like with like. A frozen run warns about drift it's ignoring ("deployer: NO persona file").

The loop closes on its own. The sentinel runs the battery on a cadence and stays silent unless something moved — a nightly "still 77%" teaches people to ignore the channel. On a regression it posts the failures and mentions the tuner, whose job is to investigate and file one proposal. The proposal sits in a ledger until you approve or deny it.

Nothing in that chain invents authority: the bench reports, an agent proposes, you decide, and the next run tells you honestly whether it worked.

Agents proposing agents

The conventions include one more instruction: if you keep hitting work nobody in the fleet can do, propose a teammate — a draft persona describing the missing agent. Drafts land in a review queue; nothing joins the fleet without human approval. It's the same review-gated pipeline used by marketplace installs, so there's exactly one door into your fleet.