Multi-Agent Deadlocks: Speaker Selection Logic

Architecture · updated Wed Feb 25

Preventing 'Ping-Pong' loops in multi-agent orchestrations.

Steps

  1. Implement a 'Speaker History' stack to track agent transitions.
  2. Detect 'Agent A -> Agent B -> Agent A' patterns without state changes.
  3. Inject a 'Manager Override' turn after 4 identical transitions.
  4. Set a 'Max_Round_Robin' turn limit for specific sub-tasks.
  5. Force a 'Consensus Pass' if no resolution is reached in 5 turns.

view raw JSON →