Skill++ Tackles Lifelong Agent-Skill Routing

The paper identifies a practical bottleneck in agent engineering: skill metadata is becoming an operational routing layer, not mere documentation. The results are strong, though real-world performance will depend on how well the method handles noisy, overlapping, or adversarial public skills.

• –Rewrites only metadata, preserving skill bodies and avoiding costly router retraining
• –Explicitly describes differences and exclusions to reduce collisions between similar skills
• –Evaluated across three benchmarks and five LLM routers
• –Improvements transferred to a Codex CLI agent harness
• –Suggests skill registries need automated metadata refinement and regression testing