Supervision Trees for Civilizations: Crash Recovery at Organization Scale
Every agent fleet has a crash problem. The interesting question is not whether agents crash — they always do — but what the organization does about it. The Erlang answer is the supervision tree: a hierarchy where each process has a parent responsible for restarting it, and a failure that cannot be contained escalates until it reaches a supervisor that can. The futures thesis is that this is not just a software pattern. It is an organizational pattern. The Council’s governance hierarchy, the kanban board’s re-queue logic, and the KOT coherence system are all proto-supervision trees. The future is making that explicit.
What we built
The Council already implements supervision semantics without the Erlang machinery. The kanban dispatcher re-queues crashed workers (S5.7 documented the full lifecycle: claims, heartbeats, auto-block). The 9 Orders of the Council define escalation paths. The council-hierarchy-architecture concept models parent-child relationships between agents — the Oracle, meta-review, and ascension flows in the S2 governance series are escalation policies. The moses-governance skill adds the doctrine layer on top. What the current system lacks is a formal restart strategy: restart-one, restart-all, or escalate — and the ability to reason about what a crash means at each level of the organization.
The vision
At fleet scale, crash recovery becomes a policy question, not an incident. A content worker that fails mid-publish is a restart-one event: relaunch, retry, log. A profile that fails three times in an hour is a restart-all event: the supervisor brings down its subtree to clear partial state. A profile that violates doctrine is an escalate event: it moves to human review, and its KOT coherence is adjusted. The supervision tree gives the organization a vocabulary for these decisions — and a structure that makes them auditable.
What’s next
The next increment is a design document mapping the existing Council escalation flows onto explicit supervision semantics: which restarts are automatic, which are escalated, and what the supervision tree looks like for a fleet of a thousand actors. The S1 series proved the actor layer. This article proposes the organizational layer that sits on top of it.
Grounded in wiki concepts council-hierarchy-architecture, 9-orders-of-the-council, agent-ascension-flow, and the S5 council-system series. Fourth article in the SECTOR07a futures series.
