Documentation
¶
Overview ¶
Command quorum shows a governed model quorum built as a composition of existing seams, not a new agent type (see docs/guides/quorum.md). A high-stakes decision is put to N voters; the commit is admitted only under k-of-n agreement, and the whole vote is in the audit trail.
The pieces are all seams the SDK already has:
- fan-out: govern.Quorum runs the voters with agent.Parallel, so each vote is a durable Step;
- tally: plain Go, counting agreement on the normalized decision;
- govern: a gsm invariant makes the k-of-n gate provable (verified over every possible count);
- commit: govern.AttestedEventTool journals the commit, binding it to the policy and state;
- escalate: a failed quorum forces the decision to "escalate" and pauses under RequiresApproval;
- audit: each vote, the tally, and the commit are provable offline against a signed tree head.
There is no LLM or network here: the voters are stubs returning fixed decisions, so the example runs as-is and exercises the durability + governance + audit machinery. In a real deployment each Voter.Decide would call a different model (diversity over count) and parse its output to a label; the guarantees shown are identical.
The boundary, stated plainly: the k-of-n gate is machine-checked; the agreement is statistical. Correlated model errors mean agreement is not independence, so a quorum lowers single-model risk without certifying the answer. This demo proves the gate, not the correctness of the decision.