Self-Improving Agent

基于对经验的持续学习，不断优化 Agent 能力。适用于完成重要任务后、出现错误时、会话结束时，或用户输入“自我进化”“总结经验”“从经验中学习”等指令时触发。

This skill appears coherent with its purpose, but exercise caution before approving any changes it proposes. Practical steps: - Review proposed skill modifications and memory appends carefully (diffs) before approving. - Back up MEMORY.md and other workspace files so you can revert unwanted changes. - Avoid storing secrets or sensitive data in workspace files the agent will read (MEMORY.md, other project files) or move such secrets out of the workspace. - Consider testing the skill in an isolated or dummy workspace first so you can observe its proposals without risk to production data. - If the skill proposes code or skill-file changes, require a manual code review or run tests before applying them. - If you need stricter guarantees, prefer a read-only evaluation run (have the agent produce proposed edits as patches) and only apply edits via a human-reviewed merge process.

Type: OpenClaw Skill Name: self-improving-agent-skill Version: 0.2.0 The skill implements a structured 'Self-Improving Agent' framework designed to learn from task experiences and optimize its own capabilities. While it possesses the high-risk capability of modifying its own instructions (SKILL.md) and the agent's core memory (MEMORY.md), it incorporates robust safety guardrails, most notably a mandatory 'User Confirmation Gate' that requires explicit human approval before any file modifications are applied. The logic is transparent, research-oriented, and includes features like evolution markers for traceability and confidence tracking to mitigate the risk of adopting poor patterns. No evidence of data exfiltration, unauthorized execution, or malicious intent was found across SKILL.md, README.md, or the associated templates.