SkillOpt: Executive Strategy for Self-Evolving Agent Skills

2026

Open publication workspace · Sign in to read the full PDF.

AI-generated summary

SKILLOPT introduces a novel text-space optimization approach to systematically enhance agent skills, leading to significant performance improvements without model weight updates.

* **Controllable Skill Optimization:** SKILLOPT acts as a text-space optimizer, turning scored rollouts into bounded edits for a single skill document, ensuring strict improvement on a validation score.
* **Robust Training Mechanism:** It employs a textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta updates for stable skill training, adding zero inference-time cost at deployment.
* **Broad Applicability and Transferability:** SKILLOPT achieves state-of-the-art results across six benchmarks, seven models, and three execution harnesses, with optimized skills demonstrating value across model scales and environments.

Tags: agent skills, LLM agents, optimization, skill evolution, text-space learning

Check the original publication for accuracy and context.