Multi-Agent Failure Modes: Collusion, Loops, and Amplified Errors
Multi-Agent Failure Modes: Collusion, Loops, and Amplified Errors
Collusion
Agents can converge on the same wrong idea because they share prompts or memory. Mitigate by injecting diversity: different prompts, different tools, different perspectives.
Looping
Agents can bounce tasks back and forth. Fix with ownership rules and max handoffs.
Amplified hallucinations
If one hallucination becomes shared state, the whole team is poisoned. Validate key facts via tools or human review.
