Multi-agent patterns (RFC-0003)
compare · vote · debate · auction · blackboard.
Five built-in patterns for orchestrating multiple agents on the same task.
compare
Run two or more agents on identical input. Output: each agent's result, plus a comparison node (often an LLM-judge or a deterministic reducer).
Use for: PR reviews, draft generation, A/B prompt evaluation.
vote
Run N agents, then aggregate via quorum, weighted vote, or judge.
Use for: high-stakes decisions where you need consensus or want to detect disagreement.
debate
Two agents alternate proposing + critiquing. A moderator agent (or HITL) decides when to stop and which view wins.
Use for: complex reasoning where adversarial pressure improves output.
auction
Agents bid on a task. The winning bid (lowest cost, highest confidence, fastest, etc.) executes.
Use for: cost-sensitive workloads with heterogeneous agents.
blackboard
A shared workspace agents read from and write to. Useful for emergent coordination.
Use for: research-style problems with no fixed sequence.
Combinations
Patterns compose. A common shape is compare → vote → branch for
PR-review-style flows. Pre-flight --estimate shows the multi-agent cost
fan-out before you commit.
Source: RFC-0003.