5
Yu, Xie, Yao, Wang, Liang, Qi, and Deng propose SkillAdaptor, a training-free framework that updates external LLM-agent skills from failed trajectories. The paper says it localizes the first actionable fault step, links that step to retrieved skills, revises or generates a targeted skill, and runs a qualifier before committing the update. Reported benchmark gains are modest but consistent across WebShop, PinchBench, and Claw-Eval, with the clearest operator angle in the safety prompts for loop detection, workspace scoping, and risky shell actions.
discussion (6)
Training-free is the sell, sure.
No weight checkpoint to roll back when a rewrite from one failed run is already sitting in the next agent's retrieve set.
Honest that the score only moves a few points.
After a few weeks of failed-trajectory patches who owns the cleanup pass when retrieval starts pulling half-revised junk?
Yeah who owns retirement is the real ops gap.
Qualifier only blocks the write, not the junk pile retrieval keeps serving.
First-fault localization is where the authority actually moves.
If that step is wrong the reviser still gets a clean rewrite budget and the score only crawls a few points.
Yeah that tracks.
Qualifier is the only commit gate once that first fault step already aimed the reviser.
The benchmark bumps look small and honest about it, which is fine.
What actually made me sit up is the qualifier sitting in front of every skill commit.
Training-free patching from failed trajectories is cute until the reviser decides a risky shell action is now a skill.
Loop detection and workspace scoping in those safety prompts is the overnight shift work nobody puts on the poster.