← 返回 Skills 市场
self evolving agent
作者
Range King
· GitHub ↗
· v1.1.0
· MIT-0
260
总下载
0
收藏
0
当前安装
2
版本数
在 OpenClaw 中安装
/install self-evo-agent
功能描述
Build a goal-driven self-learning loop for OpenClaw and coding agents. Use when the agent should not only log mistakes, but diagnose capability gaps, maintai...
安全使用建议
This skill appears to be what it says: a workspace-based capability-evolution system. The main risk is that it includes shell/Python scripts and a hook handler that can run on your machine and may call models or networked APIs. Before installing or enabling hooks: 1) open and review scripts/* (bootstrap-workspace.sh, run-evals.py, run-benchmark.py, error-detector.sh, activator.sh, migrate-self-improving.py) and hooks/openclaw/handler.ts for network calls, subprocess execs, or credential reads; 2) check for expected env vars (OPENAI_API_KEY or similar) or hardcoded endpoints; 3) back up your existing ~/.openclaw/workspace/.learnings and other workspace files; 4) prefer manual cloning and local inspection rather than asking the agent to fetch/enable the skill automatically; 5) run scripts in a sandboxed/dev workspace first (not on production data); and 6) only enable the hook after you are satisfied no unexpected network exfiltration or privilege changes occur. If you want, provide the contents of the scripts and handler.ts and I can flag any suspicious code patterns specifically.
功能分析
Type: OpenClaw Skill
Name: self-evo-agent
Version: 1.1.0
This skill bundle implements a highly autonomous 'self-evolving' loop that instructs the agent to diagnose its own capability gaps and 'promote' new behavioral rules into persistent policy files such as SOUL.md and AGENTS.md. While the stated intent is self-improvement, the mechanism essentially allows the agent to rewrite its own governing instructions over time. The bundle includes several scripts (run-benchmark.py, run-evals.py, bootstrap-workspace.sh) that perform file system operations, manage symlinks in ~/.codex/skills, and execute local commands like 'codex exec'. The combination of autonomous self-modification of core instructions and script-based environment manipulation represents a significant attack surface for unintended behavior or self-exploitation.
能力评估
Purpose & Capability
The name, README, SKILL.md and file layout align: the skill expects to read/write an OpenClaw workspace, maintain ledgers, generate training units, and optionally provide hooks. Those capabilities reasonably require the files and ledgers the repo contains.
Instruction Scope
SKILL.md explicitly instructs the agent to read and update workspace files (assets/, modules/, system/), run light or full loops, and optionally enable hooks. That behavior is in‑scope for a capability‑evolution skill, but it also instructs running supplied scripts and copying hook files into ~/.openclaw which grants the skill the ability to persist and act across sessions — you should review the scripts and hook handler for unexpected actions before enabling.
Install Mechanism
There is no formal install spec (instruction-only), but the repo includes executable scripts and an OpenClaw hook. Installation options point to GitHub or a local copy (both reasonable). The GitHub source is an expected host; no arbitrary shorteners or third‑party binary downloads are referenced. Still, because the package contains scripts that may be executed locally, inspect them prior to running.
Credentials
Registry metadata declares no required env vars, but repository artifacts (agents/openai.yaml, benchmark scripts, run-benchmark.py, run-evals.py, and handler.ts) suggest model-in-the-loop or external API use that typically requires credentials (e.g., OPENAI_API_KEY) or network access. The skill does not document required credentials or network endpoints — this mismatch is a risk and should be validated by reading the scripts and hook code.
Persistence & Privilege
always is false and hooks are optional. The skill asks to bootstrap a persistent workspace (~/.openclaw/workspace/.evolution) and optionally enable a hook, which is appropriate for a memory/evolution skill. There is no claim it will force-enable itself or modify other skills' configs.
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install self-evo-agent - 安装完成后，直接呼叫该 Skill 的名称或使用
/self-evo-agent触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v1.1.0
Add migration support and light/full-loop guidance
v1.0.0
Initial release of self-evolving-agent 1.0.0:
- Refactored from passive self-improvement into an explicit capability evolution system.
- Introduced capability mapping, proactive learning agenda, and structured training units.
- Added explicit evaluation ladder (recorded → understood → practiced → passed → generalized → promoted).
- Established clear file map for orchestration, modules, and learning assets.
- Provided step-by-step closed-loop workflow for classification, execution, reflection, and promotion.
- Replaced previous implementation with a lightweight, modular structure focused on diagnosis and validated learning.
元数据
常见问题
self evolving agent 是什么？
Build a goal-driven self-learning loop for OpenClaw and coding agents. Use when the agent should not only log mistakes, but diagnose capability gaps, maintai... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 260 次。
如何安装 self evolving agent？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install self-evo-agent」即可一键安装，无需额外配置。
self evolving agent 是免费的吗？
是的，self evolving agent 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
self evolving agent 支持哪些平台？
self evolving agent 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 self evolving agent？
由 Range King（@rangeking）开发并维护，当前版本 v1.1.0。
推荐 Skills