SkillAdaptor: Pinpoint Which Skill Broke, Fix It, Leave Everything Else Alone

The blunt approach to agent skill failures is to update the skill whenever a trajectory fails. The problem: trajectories are long, failures happen somewhere in the middle, and a full-trajectory update smears the fix across steps that were actually fine. SkillAdaptor, a new paper from Ant Group (arXiv 2606.01311), does something more precise—it identifies the first actionable fault step in a failed trajectory, traces responsibility back to the specific candidate skills, and applies targeted updates with explicit acceptance checks. The backbone model stays frozen throughout.

Tested across WebShop, PinchBench, and Claw-Eval with Kimi-K2.5, GLM-5, and GPT-5.2—three very different models—it shows consistent improvements where coarse skill update methods tend to overfit or overcorrect: +1.5 points on PinchBench Avg Score%, +1.8 on Claw-Eval. Those numbers are incremental, but they compound. Any production agent system that accumulates skills over months will drift unless updates are precise.

The framework is training-free and designed to plug into OpenClaw-class agent harnesses, which makes it immediately practical rather than a research prototype. Paper: https://arxiv.org/abs/2606.01311