SkillOpt: trajectory-driven skill optimization for frozen LLM agents

I came across Microsoft’s SkillOpt repo, which focuses on a practical problem: you have an agent whose underlying LLM is fixed, but you still want better step-by-step behavior. SkillOpt frames that as “text-space optimization,” where the system learns improved natural-language skills (prompts/instructions) rather than updating model weights.

Concretely, the workflow described in the excerpt is built around three ideas: (1) trajectory-driven edits (optimizing based on the sequences of actions/outputs you observe), (2) validation-gated updates (only accepting changes that pass checks), and (3) producing a deployable artifact, best_skill.md, meant to be carried into your real agent setup.

A situation I’d personally hit: I’m maintaining a small production RAG/agent system on top of a fixed LLM (e.g., because of latency/cost/regression constraints). I can’t fine-tune, but I do need to tighten reliability for one subtask—say, “generate search queries that match my retriever schema” or “format tool calls for a narrow tool contract.” SkillOpt’s output artifact suggests an engineer workflow where you iterate on the natural-language skill, validate it, then drop in the best version without touching the agent runtime.

What I’d look at first in the repository: the definition of the skill artifact (best_skill.md), the loop that performs trajectory-driven edits, and the validation gate that decides which candidate skills survive. If you’re already doing prompt/skill iteration manually, this is a structured way to make that process more deterministic and reviewable for an agent team.

- Artifact-first output (best_skill.md) for easy deployment - Trajectory-driven edits to guide text optimization - Validation-gated updates to prevent regressions