SkillOpt: Executive Strategy for Self-Evolving Agent Skills (arxiv.org)

The recent announcement of SkillOpt marks a significant advancement in the optimization of agent skills for AI systems. Unlike traditional methods that rely on hand-crafted or loosely controlled self-revisions, SkillOpt introduces a systematic controllable text-space optimizer that enhances skills through a disciplined training approach. This model employs a separate optimizer to transform scored rollouts into bounded edits on a skill document, ensuring that changes are only accepted when they demonstrably improve validation scores. This process is supported by innovative mechanisms like a textual learning-rate budget and a rejected-edit buffer, enabling stable training without increasing inference-time model calls. The implications of SkillOpt are profound for the AI/ML community, as it outperforms existing competitors in various contexts, achieving better or equivalent results across all evaluated benchmarks and setups. In particular, it delivers enhancements in average accuracy—up to +24.8 points in Codex agentic loops and +19.1 in Claude Code operations—significantly boosting the performance of models like GPT-5.5. Furthermore, the optimized skill artifacts demonstrate their utility across different model scales and execution environments, highlighting the versatility and potential for broader applications in AI development. This breakthrough not only improves the skill training process but also promises to simplify and elevate the design of intelligent agents.