SkillEvolver: Meta-Skill for Online Agent Learning
A new paper on arXiv (2605.10500) introduces SkillEvolver, a lightweight system enabling agents to iteratively improve domain-specific skills through real-world deployment. Unlike static skills authored once, SkillEvolver treats skill learning as a meta-skill that authors, deploys, and refines skills based on failures encountered by other agents. The system operates on skill prose and code rather than model weights, allowing seamless integration into any agent without retraining. Refinement occurs after deployment, using failure signals from actual use rather than exploratory traces. The meta-skill itself is a standard skill loadable by any protocol-compliant CLI agent.
Key facts
- SkillEvolver is introduced as a plug-and-play solution for online skill learning.
- It treats skill learning as a meta-skill that iteratively authors, deploys, and refines domain-specific skills.
- The learning target is skill prose and code, not model weights.
- Refinement occurs after deployment, using failures from other agents.
- The meta-skill is itself a skill loadable by any protocol-compliant CLI agent.
- The paper is from arXiv with ID 2605.10500.
- SkillEvolver contrasts with trace-distillation methods.
- The system is designed for any agent without retraining.
Entities
Institutions
- arXiv