Repolode
Submit
sentient-agi

sentient-agi/EvoSkill

An open-source framework that automatically creates and refines reusable skills and prompts for coding agents by learning from failed attempts.

View on GitHub
Stars
1,291
Forks
133
Open issues
21
Last commit
13 days ago

What is it?

EvoSkill is a toolkit for improving AI coding agents without manually writing their instructions. It runs an agent against a set of questions or containerized tasks, has a proposer analyze the failures, and generates new skill folders or system prompt changes to address them. Each candidate program is scored on a held-out validation set, and the best variants are kept on a frontier tracked as git branches.

The project works with several agent harnesses, including Claude Code, Codex CLI, OpenCode, OpenHands, Goose, and Harbor. Users can choose from multiple model providers such as Anthropic, OpenAI, OpenRouter, and Fireworks. Scoring can be exact match, numeric tolerance, an LLM judge, a custom script, or Harbor's built-in verifiers.

The workflow is driven by a CLI: `evoskill init` sets up a project, `evoskill run` runs the improvement loop, and commands such as `evoskill eval`, `evoskill skills`, and `evoskill diff` inspect the results. A Python API is also provided. The project is based on a published paper and is distributed under the Apache 2.0 license.

Key features

  • 01Generates new skill folders and system prompt variants from failed agent runs.
  • 02Scores candidates on a held-out validation set and keeps a frontier of top programs.
  • 03Stores each program version on a git branch for easy inspection and rollback.
  • 04Works with Claude Code, Codex CLI, OpenCode, OpenHands, Goose, and Harbor.
  • 05Supports CSV question datasets and Harbor containerized benchmark tasks.
  • 06Runs locally, in Docker, or on a Daytona sandbox.

Use cases

  • Improving a coding agent's accuracy on a specific set of question-answer tasks.
  • Evolving agent skills against containerized benchmarks such as SWE-bench through Harbor.
  • Researching automated skill discovery and agent self-improvement.

Quick start

  1. 1

    Install the project with uv (run from the cloned repo)

    uv sync
  2. 2

    Initialize a project in a git repository

    evoskill init
  3. 3

    Describe your task in .evoskill/task.md, then run the loop

    evoskill run

Things to know

  • Requires an API key for a model provider such as Anthropic, OpenAI, or OpenRouter.
  • Some agent harnesses need separately installed CLIs and specific minimum versions.
  • Continuous evolution and evolution without a benchmark are still under development.

In collections

Claude Code tools, skills & pluginsAI coding assistants & agentsAI agent skill libraries

Topics

#coding-agent#agent-skills#prompt-optimization#llm#claude-code#self-improvement#evaluation

FAQ

What is EvoSkill?
An open-source framework that automatically creates and refines reusable skills and prompts for coding agents by learning from failed attempts. EvoSkill is a toolkit for improving AI coding agents without manually writing their instructions. It runs an agent against a set of questions or containerized tasks, has a proposer analyze the failures, and generates new skill folders or system prompt changes to address them. Each candidate program is scored on a held-out validation set, and the best variants are kept on a frontier tracked as git branches.
Is EvoSkill free to use? What license does it have?
EvoSkill is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is EvoSkill still maintained?
Actively maintained. The last commit to EvoSkill was on Oct 26, 2026.
What can I use EvoSkill for?
Improving a coding agent's accuracy on a specific set of question-answer tasks. Evolving agent skills against containerized benchmarks such as SWE-bench through Harbor. Researching automated skill discovery and agent self-improvement.
How do I install EvoSkill?
Install the project with uv (run from the cloned repo): uv sync Initialize a project in a git repository: evoskill init Describe your task in .evoskill/task.md, then run the loop: evoskill run
Where can I get EvoSkill?
The source code and installation instructions are on GitHub: https://github.com/sentient-agi/EvoSkill

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects