SkillAA routes failed AI tasks to one broken skill

A paper posted September 20 targets a specific failure mode in AI agents built from skill libraries. Most repair methods edit the library after a failed run with no structured way to trace that failure back to the skill that caused it. The SkillAA paper proposes a graph that represents how skills apply to tasks, how they execute, and how they combine.

The method compares successful and failed executions on that graph and routes the repair to the specific node responsible. It updates only that local structure and leaves the rest of the library alone. Two checks, called Local and Big Gates, screen each change before it's kept.

Using gpt-5.6-sol as the base model, SkillAA reaches 81.5% on SearchQA, 66.7% on LiveMath, and 91.2% on DocVQA. None of that comes from updating model parameters. The gains are entirely from editing the skill graph.

The paper doesn't say how SkillAA compares against simpler repair baselines on the same benchmarks, and it doesn't say what the gating step costs in compute per failed rollout. Those gaps matter for anyone deciding whether the extra structure is worth building.

What the paper does establish is that routing plus gating beats routing alone would, at least on these three benchmarks. Whether that gate cost is worth paying is the open question the numbers don't answer.