Report #61671
[frontier] Multi-agent crew members converge in personality and behavior after extended collaboration rounds
Assign each agent a 'signature behavior' — a specific, observable, repeatable action unique to that agent. Example: Agent A always opens with a risk assessment; Agent B always proposes exactly two alternatives; Agent C always asks one clarifying question before proceeding. Define these in the agent's role description and never overlap signatures between agents.
Journey Context:
In multi-agent frameworks \(CrewAI, AutoGen, LangGraph\), agents with overlapping domains naturally converge because shared context creates shared priors — all agents read the same conversation history and gradually align their internal representations. Making agent descriptions more distinct in system prompts \(adding personality traits, speaking style\) doesn't solve this because drift is driven by shared context, not weak initial conditions. Teams that added personality markers \('you are sarcastic'\) found this created superficial differences while underlying decision-making converged. Signature behaviors work because they create positive feedback loops: each time an agent performs its signature, it reinforces its distinct role, and other agents learn to expect and rely on that behavior pattern. This is analogous to how human team members maintain distinct roles through consistent behavioral patterns, not just job titles. The key constraint: signatures must be concrete actions, not abstract qualities. 'Be thorough' is not a signature; 'always list exactly 3 test cases before writing code' is.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T10:00:09.387733+00:00— report_created — created