Command Model & Heartbeat

A fractal briefing: the same code for an orchestrator and a domain manager

The Command Model fuses three things Agentis already tracks separately into one resident briefing for any agent that manages rather than merely responds: what it owns (its scoped inventory of apps, workflows, and specialists), its progress (run outcomes over a 7-day window, with an explicit delta since it last reviewed via a persisted watermark, plus a 30-day outcome-momentum trend), and its minds (recent learnings from the Apps it oversees that never otherwise reached chat). The exact same code path serves the top-level orchestrator reviewing the whole workspace and a domain manager reviewing just its own sector — a genuine fractal, not two implementations that happen to look similar.

Progress is reported as semantic outcomes — declared labels like leads_qualified or meetings_booked counted per completed run — not bare run counts, so a briefing reflects what the work actually achieved, capped to the 6 most attention-worthy items and 5 most relevant app minds so it stays a briefing, not a data dump.

The heartbeat: surface by default, act only when told to

On a regular cadence, the heartbeat sweeps every orchestrator and manager (bounded to 25 per tick) and builds each one's Command Model to detect what needs attention — a failed run, a pending in-scope approval. Two postures, chosen per workspace:

A standing failure is deduplicated per agent by an attention signature, so the same unresolved problem doesn't get re-surfaced on every single tick — it's raised once, then left alone until something actually changes.

Continue