Agent Self-Improvement
Project
Agent Self-Improvement is a portable agent skill for turning real conversation history into better agent behavior. It is designed for SKILL.md-compatible agents, including Codex, Claude Code, Gemini CLI, and generic agents that support filesystem-installed skills.
Why It Exists
Most agent self-improvement drifts into vague advice or hand-picked examples. This project keeps the workflow grounded in actual conversation history: it extracts compact cards from agent sessions, lets low-cost read-only scout agents scan many shards, and keeps final judgment with the main agent.
What It Does
- Reads Codex sessions, Claude projects, Gemini exports, and generic transcript folders.
- Produces compact conversation-card shards plus a manifest for verification.
- Uses role-based scouts for broad coverage and a main synthesizer for final rules.
- Defaults to scope-safe updates for the agent itself; project, memory, and global findings stay report-only unless explicitly requested.
- Includes an installer for Codex, Claude Code, Gemini CLI, generic agents, and custom skill directories.
Install
Give this GitHub link to a local coding agent:
Install this skill: https://github.com/Chenwei-1999/agent-self-improvement