A new research framework lets AI agents repair their own broken skills without retraining the underlying model or guessing which part needs fixing.
The system, called SkillAA (Skill Abductive Attribution), stores an agent's skills, the step-by-step procedures it uses to complete tasks, in a single graph. When a task fails, SkillAA compares that failed run against a successful one and traces the error to a specific node in the graph, instead of editing skills blindly from the failure alone. It updates only that local piece of the graph, then runs the proposed change through two checks the researchers call Local and Big Gates before letting it stick. Tested with a model called gpt-5.6-sol, the approach scored 81.5% on a search-based question answering benchmark, 66.7% on a live math benchmark, and 91.2% on a document visual question answering benchmark, topping every other method in each setting.
Most agent skill libraries today work like a junk drawer. Instructions pile up, get edited whenever something breaks, and nobody tracks exactly what caused the failure or whether the fix broke something else. SkillAA's structured approach, isolate the fault, patch only that piece, validate before committing, looks a lot more like how a careful engineer handles a bug report than how most current agent tooling operates.
The results are strong but the test set is narrow, three benchmarks and one base model, so it is too early to say whether attribution-guided graph editing scales to messier, real-world agent work.