SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent
SkillProx evolves reusable agent instructions through measured retries, regression rollback, and utility-based pruning, reporting a 3-point accuracy gain over its strongest baseline.
SkillProx treats agent skills as editable text rather than model weights. Its **forward stage** reruns diagnosis-driven edits on the same task batch, rolls back regressions, and returns measured outcomes; its **backward stage** audits individual knowledge units for consolidation, demotion, or removal.
For recurring coding-agent work, the useful pattern is closed-loop skill maintenance: test each instruction change, preserve the prior version when results worsen, and periodically remove guidance that contributes no measured utility. This makes accumulated procedures more auditable than an ever-growing prompt file.
SkillProx treats agent skills as editable text rather than model weights. Its **forward stage** reruns diagnosis-driven edits on the same task batch, rolls back regressions, and returns measured outcomes; its **backward stage** audits individual knowledge units for consolidation, demotion, or removal. For recurring coding-agent work, the useful pattern is closed-loop skill maintenance: test each instruction change, preserve the prior version when results worsen, and periodically remove guidance that contributes no measured utility. This makes accumulated procedures more auditable than an ever-growing prompt file. Across multiple backbones and in- and out-of-distribution benchmarks, SkillProx reports an average **3.0 percentage-point** accuracy gain over the strongest gradient-based baseline. The abstract does not identify the tasks, absolute scores, audit cost, or whether the approach transfers to live software repositories.