Skill‑policy co‑evolution denotes the mutual improvement of an agent's policy model and its skill bank during training, a promising paradigm for self‑improving agents. Existing evaluations typically score the co‑evolved policy on the final skill bank it trained with, the so‑called delivered bank.
After deployment, however, the skill bank may be rewritten, added, or removed far more often than the policy is retrained, forcing the policy to operate on a drifted bank that neither the training objective nor current evaluation protocols address.
Theoretical analysis shows this blind spot is systematic. The cooperative acceptance rule shared by co‑evolution methods ties a policy's competence to the bank it co‑evolved with, making it fragile once skills drift; empirically, sensitivity to drift grows as co‑evolution training proceeds.
We therefore introduce the CO‑evolution Drift Audit (CODA), a deployment‑oriented evaluation protocol that simulates skill drift. CODA measures worst‑case success rate and adds two metrics: a consistency rate under meaning‑preserving edits and a resistance rate to corrupted skills.
Across four co‑evolution methods, five benchmarks and six settings, worst‑case success falls 5‑13 points below the peak measured for each delivered pair, erasing 52‑93% of the delivered bank's gain over the no‑skill baseline.
CODA reframes the co‑evolution target from peak performance on the co‑evolved bank to the ability to harness skills as they evolve, providing a shared standard for future audits.
Review