Preprint: Self-Improving Agents Lose Half Their Gains When Skill Banks Drift Post-Deployment

Self-improving AI agents often learn by co-evolving a policy model with a library of skills. Current standard evaluations test the agent against the exact skill bank it was trained on, assuming this represents its capability. However, in real-world deployment, skills are frequently added, removed, or rewritten without retraining the underlying policy model.

A new preprint introduces the CO-evolution Drift Audit (CODA), an evaluation protocol that simulates this post-deployment skill drift. The researchers tested four different co-evolution methods across five benchmarks. They found that under worst-case drift conditions, success rates fell 5 to 13 percentage points below the peak scores measured on the original delivered bank.

This drop erased between 52% and 93% of the performance gains that the co-evolved skill bank provided over a baseline with no skills at all. The study argues that the shared acceptance rule in these methods ties competence too tightly to the specific training bank, making agents fragile once that bank changes. CODA proposes new metrics for consistency under edits and resistance to corrupted skills, shifting the goal from peak performance to robust harnessing of evolving skills.