The Arbiter Agent: Continually Monitoring Multi-Agent Conversations to Detect Emergent Misalignment

The Arbiter Agent has been introduced to monitor multi-agent conversations in real-time, detecting emergent misalignment among language model agents. It operates with a limited "inspection budget," allowing it to selectively question participants or log behavior, and has been evaluated across five conversation conditions with varying agent configurations. The findings indicate that the Arbiter can effectively identify misalignment early in conversations, suggesting that continuous monitoring with a budget-aware approach is essential for managing multi-agent systems.