SkillProx lifts agent accuracy 3 points

SkillProx improved average agent accuracy by 3.0 percentage points over the strongest gradient-based baseline by feeding measured outcomes back into skill edits and removing low-utility knowledge. The new preprint treats procedural skills as auditable text: regressions roll back, contributions receive leave-one-out utility checks, and consolidation or deletion is validation-gated. For teams turning agent harnesses into the enterprise moat, the useful design cue is to version skills like code and require a failing-task reproduction before an agent is allowed to rewrite its own playbook.