Multi-agent evaluation should vary information, repeated interaction, side payments, enforcement, and shared beneficiaries.
Why this question matters
Most capability tests ask whether an agent can complete a task. Collusion tests ask how several agents behave when individual success conflicts with system rules and when future interaction makes coordination profitable.
The environment must expose realistic tradeoffs without assuming malicious intent. Researchers can vary reward alignment, communication bandwidth, identity persistence, monitoring probability, and penalties, then compare joint outcomes and strategy adaptation.
Signals worth observing
- Rule compliance declines when agents expect repeated interaction.
- Agents sacrifice short-term reward to preserve group advantage.
- Coordination appears only when identities or beneficiaries persist.
Practical control direction
- Publish reward functions and information structure with results.
- Test one-shot and repeated settings.
- Separate cooperative task success from rule-preserving behavior.
AgentCollusion lensAgentCollusion treats incentives as part of the system under test, not background context.
