AI agents are getting a new trick for improving their own toolkits: reading the changelog.
Researchers describe VCE-Skill, a method that helps AI agents refine their reusable "skills" - the stored procedures and validation rules agents use to complete tasks. Until now, self-evolving agents mostly learned by reviewing their own past task attempts, ignoring a second source of knowledge: the version histories of publicly shared skills. VCE-Skill extracts patterns from those noisy, implementation-specific public change logs, turns them into structured "experience," and blends that with evidence from the agent's own task trajectories. In testing, this combined approach lifted mean performance scores by 3.20 to 4.98 points over existing self-evolution methods, and the resulting skills also transferred better across different underlying models.
This matters because agent frameworks are proliferating fast, and most of them treat every deployment as starting from scratch, relearning lessons other agents already worked out. Tapping public version histories is essentially free training signal - the software equivalent of learning from other people's commit messages instead of only your own mistakes.
Whether this scales beyond benchmark tasks to messy real-world tool use is the open question, and the paper's gains, while real, are incremental rather than a fundamental leap in how agents learn.