Microsoft has released SkillOpt, an open-source tool that optimizes AI agent abilities using a novel training loop. The v0.2.0 update on July 2, 2026, treats natural language skills as trainable components for frozen LLMs, improving them without altering model weights, according to the project's GitHub repository.
Traditionally, agent skills are hand-crafted or evolved through loosely controlled self-revision. These methods often fail to produce reliable improvements. SkillOpt introduces a structured, disciplined approach modeled after deep-learning optimization.
The process treats a skill document, typically a Markdown file, as the trainable state. An optimizer model analyzes agent performance on tasks and proposes edits. The final output is a compact `best_skill.md` file that can be deployed with the original, unchanged LLM, adding no inference-time overhead.
How Does SkillOpt Train Agent Skills?
SkillOpt trains skills through a deterministic loop: rollout, reflect, aggregate, select, and update. An optimizer LLM proposes edits to the skill text based on the agent's performance on a task. Crucially, a candidate edit is only accepted if it improves a score on a separate, held-out validation set.This validation-gated process prevents performance regressions, a common problem in agent self-improvement. The framework also includes concepts like a textual learning-rate budget and a rejected-edit buffer to stabilize the training process, making skill evolution more predictable and reproducible.
What Are the Performance Gains?
Testing shows significant improvements across multiple benchmarks and models. For example, SkillOpt lifted a GPT-5.5 model's average accuracy by +23.5 points in direct chat and +24.8 points within a Codex agent loop. The optimized skills are also transferable across different models and tasks.
The framework supports a variety of backends, including OpenAI, Azure, Claude, and Qwen, as well as execution harnesses for agentic frameworks like Codex and Claude Code. This makes it adaptable to diverse development environments.
Method | Process | Output | Reliability |
|---|---|---|---|
Hand-Crafting | Manual prompt engineering by a human. | Static prompt text. | Low; depends entirely on human intuition. |
Self-Revision | Agent critiques and rewrites its own prompts. | Evolving prompt text. | Variable; can lead to regressions without strict validation. |
SkillOpt | Automated, validation-gated edits based on performance. | Optimized `best_skill.md` file. | High; strictly improves or maintains validation score. |
The New SkillOpt-Sleep Feature
The v0.2.0 release introduces SkillOpt-Sleep, a nightly offline engine for self-evolution. It automatically harvests an agent's daily task history, mines recurring patterns, replays them to test new strategies, and consolidates the best results into an improved skill file behind a validation gate.This feature provides a mechanism for continuous, on-the-job training for agents. It allows a local coding assistant like Codex or Copilot to learn from its own work overnight, creating a self-improving loop. This approach is similar in spirit to other reinforcement learning frameworks like the Agent Reinforcement Trainer from OpenPipe, which also focuses on improving agents through experience.
