SkillOpt is an open-source framework from Microsoft, with more than 18,500 stars on GitHub, that trains the natural-language skills of LLM agents the way neural networks are trained โ with epochs, batches, learning rates, and validation gates โ without touching any model weights. Instead of hand-crafting agent skills or letting agents revise themselves unchecked, SkillOpt treats a skill document as trainable state and optimizes it against held-out validation scores. The project, including its SkillOpt-Sleep nightly self-evolution engine, is under active development in 2026.
SkillOpt runs a disciplined loop: rollouts of an agent using the current skill are scored, an optimizer model reflects on the results and proposes bounded add, delete, or replace edits to the skill text, and a candidate edit is accepted only if it strictly improves a held-out validation score. A textual learning-rate budget and a rejected-edit buffer keep training stable, and the final artifact is a compact best_skill.md file of a few hundred to two thousand tokens that deploys with zero extra inference-time model calls. Microsoft reports best or tied-best results across six benchmarks, seven target models, and three execution harnesses, with large accuracy gains over no-skill baselines.
SkillOpt is completely free and open source under the MIT license, installable from PyPI. You pay only for the model API calls used during training and evaluation.
SkillOpt is for agent builders and research teams who rely on reusable skills and want those skills to improve measurably and reproducibly instead of drifting with every hand edit.