A team led by Jinluan Yang posted a new arXiv paper on Oct 16, 2026. It shows that agents trained to improve both their policy and a skill bank become fragile if the skill bank changes later. The authors propose a new evaluation called CO‑evolution Drift Audit (CODA) that simulates skill drift and measures consistency and resistance. Across four methods and five benchmarks, CODA finds worst‑case success drops 5‑13 points, wiping out most of the advantage the agents had over a no‑skill baseline.
Why it matters
Deployers will need to test AI agents for robustness to skill changes, not just peak performance.
