self-improving-agent-python

Implement a 3-layer self-improvement process for agents to evaluate tasks, learn from outcomes, optimize performance, and share knowledge across agents.

Security Scan

The skill's code and runtime instructions are consistent with its stated purpose (local evaluation, lesson recording, optimization and cross-agent file-based syncing) and it does not request credentials or network access, but it will read/write JSON files across all workspaces under ~/.workbuddy which you should be aware of.

Purpose & Capability

Name/description describe a local self-improvement system; the provided Python scripts implement evaluation, lesson recording, optimization, and cross-agent sync via local files under the workspace directory—this matches the stated purpose.

Instruction Scope

SKILL.md instructs running the included scripts and documents the filesystem layout. The runtime instructions and scripts only read/write local JSON files and do not attempt to access network endpoints, credentials, or unrelated system files.

Install Mechanism

There is no install spec and no external downloads. The skill is distributed as plain Python scripts (no package installation or remote fetch), which is low-risk from an install perspective.

Credentials

The skill requests no environment variables or credentials. It does assume a WorkBuddy-style workspace directory (DEFAULT_OPENCLAW_DIR = ~/.workbuddy) and will operate on files under that hierarchy; no secrets are required or requested.

Persistence & Privilege

always:false and no autonomous-model restrictions. However, sync_learning.py enumerates all workspaces under ~/.workbuddy and copies collective-wisdom.json into each workspace's shared-context/self-improvement directory — this intentionally modifies other workspace directories on the same host. This behavior matches the 'cross-agent sharing' purpose but is a persistent file-write capability that you should consider before using on multi-user or shared machines.

Assessment

This skill appears to do what it says: it stores evaluations, lessons, and optimization plans as JSON files in your WorkBuddy workspace and can copy a shared 'collective-wisdom.json' into every workspace under ~/.workbuddy. Before installing or running it: (1) confirm you are okay with the skill creating and writing files under ~/.workbuddy and across any workspace-* directories on the machine, (2) inspect the included scripts (they are all present and readable) and backup any existing shared-context/ self-improvement files you care about, (3) if you run sync_learning.py on a multi-user or shared machine, be aware it will write into other workspaces under ~/.workbuddy, and (4) run the scripts as a non-privileged user and in an isolated workspace if you want to limit impact. No network endpoints or credentials are used by the skill.

Self-Improvement Agent Skill

当 Agent 需要从任务执行中学习并自我优化时，激活此技能。

---

激活条件

用户提到以下关键词时激活：

• "自我改进"、"自我优化"、"学习经验"
• "任务评估"、"性能分析"
• "优化工作流程"、"提升效率"
• 需要 Agent 从错误中学习的场景

---

核心概念

Agent 不应该只是执行任务，还应该从执行中学习，不断优化自己的表现。

三层自我改进机制：

1. Layer 1: 实时反馈循环 - 每次任务执行后立即评估
2. Layer 2: 周期性深度反思 - 定期深度分析
3. Layer 3: 跨 Agent 经验共享 - 共享学习成果

---

任务评估标准

评分公式

总分 = 完成度(30%) + 效率(20%) + 质量(30%) + 满意度(20%)

评分等级

分数	等级	行动
≥90	优秀	记录最佳实践
80-89	良好	继续保持
70-79	及格	识别改进点
<70	不及格	触发深度反思

---

使用方法

Python 版本（推荐）

# 评估任务
python scripts/evaluate_task.py \
    --agent-id "my-agent" \
    --task-id "task-001" \
    --task-type "内容创作" \
    --completion 90 \
    --efficiency 85 \
    --quality 80 \
    --satisfaction 85

# 记录经验
python scripts/learn_lesson.py \
    --agent-id "my-agent" \
    --lesson "验证链接有效性" \
    --impact high \
    --category quality

# 优化分析
python scripts/optimize_agent.py --agent-id "my-agent"

# 跨 Agent 同步
python scripts/sync_learning.py

参数说明

evaluate_task.py

参数	说明	必填
--agent-id	Agent 标识	是
--task-id	任务标识	是
--task-type	任务类型	是
--completion	完成度 (0-100)	是
--efficiency	效率 (0-100)	是
--quality	质量 (0-100)	是
--satisfaction	满意度 (0-100)	是

learn_lesson.py

参数	说明	必填	可选值
--agent-id	Agent 标识	是
--lesson	经验描述	是
--impact	影响程度	是	high/medium/low
--category	类别	是	quality/efficiency/tools/knowledge

---

数据存储位置

{workspace}/
├── self-improvement/
│   ├── evaluations.json        # 评估记录
│   ├── lessons-learned.json   # 经验库
│   └── optimization-plan.json # 优化计划
└── shared-context/
    └── self-improvement/
        └── collective-wisdom.json  # 共享知识库

---

最佳实践

何时评估

• 完成重要任务后
• 完成复杂任务后
• 任务失败或返工后
• 简单重复性任务可选

如何记录经验

• 记录具体的失败场景
• 记录成功的关键因素
• 记录工具使用的注意事项

如何应用学习

• 定期回顾经验库
• 主动应用最佳实践
• 分享给其他 Agent

---

故障排查

问题	解决
找不到工作区	检查 config.py 中的路径配置
经验未同步	手动运行 sync_learning.py
优化计划生成失败	确保有足够的历史评估数据

---

让每个 Agent 都成为终身学习者！

Version: 1.1.0 (Python Edition) Tags: agent, self-improvement, learning, optimization