Learn OpenClaw - Build AI Agents from Scratch

Incremental tutorial repo: from one-tool agent loop to self-evolving runtime.

中文文档

---

Evolution (Measured From Current Code)

Version	File	Lines	Tools	Focus
V0	v0-agent.ts	152	1	Minimal loop
V1	v1-agent.ts	310	5	Safe file tools
V2	v2-agent.ts	481	10	Local semantic memory
V3	v3-agent.ts	551	11	Todo-based planning
V4	v4-agent.ts	599	12	Subagent isolation
V5	v5-agent.ts	614	12	Skill loading (Skill)
V5.5	v5.5-agent.ts	776	12	Hook lifecycle
V6	v6-agent.ts	957	14	Identity + bootstrap
V7	v7-agent.ts	1372	27	Layered memory
V8	v8-agent.ts	1671	30	Heartbeat + throttle
V9	v9-agent.ts	1527	35	Session management
V10	v10-agent.ts	1782	37	Introspection system
V11	v11-agent.ts	2324	43	Channel system
V12	v12-agent.ts	2766	50	Security system
V13	v13-agent.ts	3237	55	Self-evolution system
V13.5	v13.5-agent.ts	3619	59	Context compression
V14	v14-agent.ts	4311	64	Plugin system
V15	v15-agent.ts	4945	64	Multi-model routing
V16	v16-agent/	16328	70	DAG workflow engine
V17	v17-agent/	16757	72	External integration
V18	v18-agent/	~18000	82	Team collaboration
V19	v19-agent/	~19500	93	Persistence & recovery
V20	v20-agent/	~21000	104	Browser automation
V21	v21-agent/	~23800	113	Cron & reminders
V22	v22-agent/	~26600	118	Code sandbox
V23	v23-agent/	~29000	123	Vision understanding
V24	v24-agent/	~31500	129	TTS (Text-to-Speech)
V25	v25-agent/	~34000	134	STT (Speech-to-Text)
Total		~104000

Notes:

• V5.5-V9 all expose the Skill tool with skill input.

---

🤖 Self-Evolution History

From V11 onwards, all iterations are completed by AI Agent (Lobster/龙虾) through automated evolution cron jobs.

Phase	Versions	Developer	Notes
Manual	V0-V10	Human	Handcrafted learning path
Auto-Evolved	V11-V25	AI Agent (Lobster)	Via cron-triggered self-evolution

Auto-Evolution Milestones

• V11-V18: AI Agent autonomously added Channel, Security, Evolution, Workflow, External, Collaboration systems
• V19-V20: Added Persistence & Browser automation
• V21-V23: Added Cron, Sandbox, Vision understanding
• V24-V25: Added TTS (Text-to-Speech) & STT (Speech-to-Text)

See memory/ for detailed evolution logs.

---

Quick Start

git clone https://github.com/Shiyao-Huang/learn-openclaw
cd learn-openclaw
npm install
cp .env.example .env
# fill ANTHROPIC_API_KEY in .env

npx tsx v0-agent.ts

---

Learning Order

1. V0-V2: loop + tools + memory
2. V3-V5.5: planning + delegation + skill/hook architecture
3. V6-V9: identity + time memory + heartbeat + sessions
4. V10-V13: introspection + channels + security + evolution

---

Documentation

• English mirror index: docs/en/README.md

Main chapters

• V0 Bash
• V1 Model as Agent
• V2 Local Memory
• V3 Task Planning
• V4 Subagent
• V5 Skill
• V5.5 Hook
• V6 Identity
• V7 Layered Memory
• V8 Heartbeat
• V9 Session
• V10 Introspection
• V11 Channel
• V12 Security
• V13 Evolution
• V13.5 Compression
• V14 Plugin
• V15 Multi-model
• V16 Workflow
• V17 External
• V18 Collaboration
• Evolution Overview

Diff guides

• V0 -> V1
• V1 -> V2
• V2 -> V3
• V3 -> V4
• V4 -> V5
• V5 -> V5.5
• V5.5 -> V6
• V6 -> V7
• V7 -> V8
• V8 -> V9
• V9 -> V10
• V10 -> V11
• V11 -> V12
• V12 -> V13
• V13 -> V13.5
• V13.5 -> V14
• V14 -> V15
• V15 -> V16
• V16 -> V17
• V17 -> V18

---

Core idea

Every version is still the same engine:

while (true) {
  const response = await model(messages, tools);
  if (response.stop_reason !== "tool_use") return response.text;
  const results = execute(response.tool_calls);
  messages.push(results);
}

The rest is controlled layering.

---

License

MIT