SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision

Researchers introduce SkillRevise, a framework that automatically refines LLM agent skills through execution-grounded iteration, improving task success rates from 36% to 62% on benchmarks. The approach addresses the cold-start problem in agent development by diagnosing defects from execution traces and applying targeted repairs, while demonstrating strong cross-model transferability.

SkillRevise represents a meaningful advance in autonomous agent development by solving a practical bottleneck: how to bootstrap effective agent skills without expensive expert authoring or weak one-shot generation. The framework operates on a simple but powerful principle—learning from execution failures to diagnose and repair skill defects systematically. This addresses a real pain point in the AI agent ecosystem, where procedural artifacts often fail in ways that aren't apparent from static code review.

The research builds on existing self-evolving methods but introduces execution tracing as the diagnostic foundation, retrieving repair principles from a general memory rather than starting from scratch. The 71% relative improvement in success rates (36.05% to 61.63%) on SkillsBench demonstrates substantial practical gains. Critically, the cross-model transferability finding suggests that refined skills capture generalizable procedural knowledge, not model-specific artifacts—meaning improvements benefit developers across different LLM architectures.

For the AI agent market, this compounds the value proposition of autonomous systems by reducing dependency on expert skill engineering and accelerating the development cycle. As agent applications expand into enterprise and specialized domains, automated skill refinement becomes increasingly valuable. The framework's empirical grounding makes it actionable for practitioners building production agent systems, though broader adoption will depend on integration into popular agent frameworks.

The technique opens interesting research directions around skill composability and transfer learning in procedural knowledge. Developers should monitor how this integrates with emerging agent orchestration platforms and whether performance gains hold on increasingly complex, domain-specific tasks beyond current benchmarks.

• →SkillRevise achieves 71% relative improvement in agent success rates through automated execution-grounded skill refinement.
• →The framework solves the cold-start problem by diagnosing defects from execution traces and applying targeted repairs without expert intervention.
• →Refined skills demonstrate strong cross-model transferability, suggesting the method captures generalizable procedural knowledge rather than model-specific patterns.
• →Evaluation across five LLMs and three benchmarks validates the approach's robustness and broad applicability.
• →Automated skill refinement reduces dependency on costly expert authoring while outperforming weak one-shot LLM generation baselines.