TL;DR
A framework that optimizes natural language skills for frozen LLM agents by treating skill documents as trainable parameters, using trajectory-driven edits and validation gates.
Key features
Text-space optimization: Applies deep learning concepts (epochs, mini-batches, learning rates) to skill document refinement.
Trajectory-driven editing: Analyzes agent execution traces to make precise skill modifications.
Validation gating: Evaluates candidate skills on held-out validation sets, accepting updates only upon confirmed improvement.
Multi-backend support: Works with OpenAI, Azure, Claude, Qwen, and other LLM providers.
Deployable artifacts: Saves optimized skills as reusable best_skill.md files.
When to use it
When you need to automatically improve an LLM agent's task performance without manual prompt engineering.
When you want robust, iteratively refined natural language skills based on multiple execution traces.
When you seek to enhance agent capabilities without fine-tuning the underlying model weights.