sentient-agi/EvoSkill

What it solves

EvoSkill automates the process of turning general-purpose AI coding agents into high-performing specialists. It solves the problem of manually tuning prompts and creating helper scripts for specific tasks by automatically discovering, creating, and improving "skills" (reusable instructions and scripts) based on agent failures and benchmark performance.

How it works

EvoSkill implements a self-improvement loop that evolves an agent's capabilities. It proposes multiple mutations to the agent's skills and prompts jointly, evaluates these variants against a dataset (either static CSV pairs or containerized Harbor tasks), and retains the best-performing versions. Each iteration produces a new agent program. These evolved skills are packaged as reusable folders containing instructions and metadata, allowing them to be transferred across different agents, LLM models, and even different but related tasks.

Who it’s for

It is designed for developers and AI researchers who use coding agents (such as Claude Code, OpenHands, or Goose) and want to optimize their performance on specific benchmarks or real-world coding tasks without manual prompt engineering.

Highlights

• Multi-Agent Compatibility: Works with Claude Code, Codex CLI, OpenCode, OpenHands, Goose, and Harbor.
• Automated Evolution: Uses a feedback-driven loop to mutate prompts and skills based on performance data.
• Cross-Transferability: Skills can be transferred across different agents, different LLM providers, and related tasks.
• Flexible Evaluation: Supports various scoring methods including exact matching, LLM-as-judge, custom shell scripts, and Harbor's built-in container verifiers.
• Remote Execution: Supports running the evolution loop in Docker or via Daytona sandboxes to handle long-running optimization processes.