CODA Reveals a Skill-Policy Co-evolution Trap in Self-Improving Agents

Summary

Skill-policy co-evolution lets an agent's policy model and skill bank improve one another, but existing evaluations usually test only the final bank produced during training. This paper argues that such evaluations hide a deployment risk: skills can be rewritten, added, or removed more often than the policy is retrained. The authors' theoretical and empirical analysis links this fragility to cooperative acceptance, a rule shared by the studied methods that binds policy competence to its co-evolved bank. They report that sensitivity to skill drift increases as co-evolution continues. To measure the problem, they introduce the CO-evolution Drift Audit (CODA), which simulates drift and reports worst-case success, consistency under meaning-preserving edits, and resistance to corrupted skills. Across four co-evolution methods and five benchmarks in six settings, worst-case success was 5-13 percentage points below the peak for each delivered policy-bank pair. This removed 52-93% of the delivered bank's improvement over a no-skill baseline. The authors therefore propose evaluating whether agents can use skills as those skills evolve, rather than measuring only peak performance on the bank used for training.