SkillOpt: Executive Strategy for Self-Evolving Agen... | AI Research

SkillOpt: Executive Strategy for Self-Evolving Agent Skills
The researchers behind SkillOpt address a fundamental challenge in AI agent development: how to reliably improve an agent’s procedural performance without needing to retrain its underlying model weights. While current methods often rely on manual prompt engineering or uncontrolled self-revision, SkillOpt introduces a systematic, "deep-learning-style" optimizer for agent skills. By treating a skill document as an external, trainable state, the system uses a separate optimizer model to iteratively refine procedural instructions, ensuring that only changes that demonstrably improve performance on held-out data are accepted.

A Controlled Approach to Skill Evolution

SkillOpt functions like a traditional machine learning optimizer but operates entirely in text space. The process begins with a frozen target model executing tasks using a current skill document. The system then analyzes the resulting successes and failures, grouping them into reflection minibatches. An optimizer model proposes specific, bounded edits—such as adding, deleting, or replacing instructions—to the skill document. To maintain stability, these edits are subject to a "textual learning rate" and a validation gate: a candidate skill is only accepted if it improves performance on a held-out validation set. This prevents the agent from adopting harmful or overfitting changes.

Stability and Negative Feedback

A key innovation in SkillOpt is its ability to learn from its own mistakes. The system maintains a "rejected-edit buffer," which records failed attempts and the reasons for their poor performance. This buffer provides negative feedback to the optimizer, ensuring that the model does not repeat ineffective strategies in future iterations. Additionally, the system employs an "epoch-wise slow/meta update," which acts similarly to a momentum term in deep learning. This allows the system to capture long-term, stable improvements across training epochs while keeping the final, deployed skill artifact compact and easy to audit.

Proven Performance and Portability

The researchers evaluated SkillOpt across six benchmarks, seven target models, and three different execution environments, including direct chat, Codex, and Claude Code. In all 52 evaluated scenarios, SkillOpt either outperformed or tied with existing methods, including human-written skills, one-shot prompting, and other automated evolution techniques. Notably, the optimized skills are highly portable; a skill trained in one environment or on one model scale often retains its effectiveness when transferred to different models or execution harnesses. This allows developers to optimize a skill once and deploy it across various agentic systems without additional training.

Practical Implications

The final output of the SkillOpt process is a compact, human-readable file (typically 300 to 2,000 tokens) that serves as a persistent procedural memory for the agent. Because the optimization happens offline and the resulting artifact is just text, there is zero additional inference-time cost when the skill is deployed. This makes SkillOpt a practical, harness-agnostic tool for domain adaptation, enabling agents to improve their tool-use, formatting, and reasoning capabilities through a rigorous, data-driven process that does not require modifying the agent's core model weights.