Self-Improving Agent Skills
Self-Improving Agent Skills is an open-source application that automatically optimizes agent skills through a multi-agent improvement loop built on Google ADK and Gemini. Three collaborating agents — Executor, Analyst, and Mutator — run and score a skill, diagnose its failures, and apply one targeted prompt change at a time, keeping only changes that improve the score (a design described in the project's documentation; we have not tested it).
Why it caught our eye
Rather than relying on static prompt templates, the app treats skill improvement as a measurable, iterative loop: each candidate change is evaluated against test scenarios, kept only if the score improves, and otherwise reverted, repeating until a target pass rate or maximum rounds is reached. This automates a manual prompt-tuning workflow with a structured three-agent division of labor.
What it does
The app optimizes agent skills using a multi-agent system built with Google ADK and Gemini. An Executor agent runs and scores skills, an Analyst agent diagnoses failures, and a Mutator agent makes one targeted prompt change. Changes are retained when the score improves and reverted when it does not, repeating until a target pass rate or maximum number of rounds.
Intended users and input
It is aimed at people who author agent skills. Users upload a skill folder as a .zip following the agentskills.io spec, with a required SKILL.md; before optimization begins they can review and edit the generated test scenarios and binary evaluation criteria.
Output
After optimization, the app produces an improved SKILL.md with successful changes applied, a detailed changelog, and a baseline-versus-final performance comparison.
Access and setup
It runs locally, with a FastAPI backend on port 8891 and a Next.js frontend on port 3000. A Gemini API key from Google AI Studio is required and is entered in the UI rather than persisted.
Known limitations
Documented constraints include the skill input format (a .zip following the agentskills.io spec with a required SKILL.md) and the dependency on a Gemini API key. The loop is bounded by a target pass rate or maximum rounds, so it terminates rather than running indefinitely.
Sources & freshness
- raw.githubusercontent.com Checked 2026-10-23
AI-assisted introduction based on public sources. We have not conducted a hands-on review. Check the official site for current access and pricing.