poly-gents Consensus & arbitration
Multi-agent systems often have silent disagreements that surface as inconsistent outputs to the user. This skill resolves them on purpose.
When to invoke
- Two agents propose conflicting plans for the same task.
- A reviewer agent (e.g.,
shadow-code-reviewer) and the author agent disagree on whether to ship. - Two roles each claim ownership of the same artefact (a doc, a ticket, a code path).
The arbitration rule
In priority order:
- Hard policy beats opinion. If a policy or compliance constraint applies, that wins. Cite the policy explicitly.
- The role with the closest scope wins. A frontend-developer wins on UI choice; a backend-developer wins on schema choice; a security-specialist wins on threat model. If unclear, the manager (
crew-manager-pattern) decides. - Existing precedent beats novelty. If the codebase / playbook already has a pattern for this, keep it. Novelty needs explicit user buy-in.
- If still unresolved, escalate to a human. Frame the decision as a short, neutral writeup with the two options, the trade-off, and your default if no answer arrives by a deadline (see
clarification-checkpoints).
The arbitration write-up
Whichever agent is hosting the arbitration writes one message:
Decision: <one sentence>
Rule applied: <which of the rules above, or "human escalation">
Loser's concern preserved: <one sentence — what we lose by not picking the other option, so we can revisit later>
Owner: <agent or human that will execute the decision>
Anti-patterns
- Letting the conflict linger and producing inconsistent outputs.
- Two agents quietly editing the same file with conflicting intents.
- Skipping the loser's-concern note — that's what catches regressions in retrospectives.