# learn-claude-code — Build an Agent Harness

> learn-claude-code teaches agent harness engineering: tool dispatch, worktrees, context compression, and teams. Clone, set API key, run the demo scripts.

## Install

Copy the content below into your project:

## Quick Use

1. Install the verified entrypoint:

```bash
git clone https://github.com/shareAI-lab/learn-claude-code && cd learn-claude-code && pip install -r requirements.txt
```

2. Run a first local check:

```bash
cp .env.example .env && python agents/s01_agent_loop.py
```

3. Add a repeatable verification step:

```bash
python agents/s12_worktree_task_isolation.py
```

---

## Intro

learn-claude-code is a verified GitHub-backed asset sourced from `shareAI-lab/learn-claude-code` with 59,755 stars and a MIT license snapshot. Best for: engineers building agent products who need a concrete, runnable reference harness rather than high-level blog posts. Works with: Python + pip; optional web UI (npm) for visualizations. Setup time: 12 minutes.

### Quantitative notes

- Multiple runnable stages (s01 → s12, repo)
- Setup time ~12 minutes

---

## Deep Dive

### What it solves

Use this when you need a **repeatable, team-shareable workflow** instead of one-off agent prompts. The goal is to make installation, first-run validation, and rollback **predictable**.

### Minimal mental model

- Treat the GitHub repo as the source of truth: install instructions, configs, and upgrade paths live there.
- Keep your first run small: one command, one verification, one rollback plan.
- Capture a baseline: setup time, first successful run, and one real task completed end-to-end.

### Safe rollout checklist

1. **Verify source**: confirm repo URL, stars, and license match what you expect.
2. **Install** using the Quick Use commands above.
3. **Prove it works** with the verification command; save the output in a note or issue.
4. **Operationalize**: document owner, upgrade command, and rollback command.

### Troubleshooting (common)

- **Install succeeds but nothing shows up**
  - Likely cause: the tool needs a restart/reload (CLI/IDE) or a config file in the right path.
  - Fix: restart the client, then re-run the verification step.

- **Works on one machine, fails on another**
  - Likely cause: Node/Python/Docker versions differ or missing system dependencies.
  - Fix: pin versions (Node/Python), and copy a minimal known-good config.

- **Token cost or latency is worse than expected**
  - Likely cause: tool schemas or verbose outputs get injected into context.
  - Fix: prefer smaller steps, cache results, and keep tool responses concise when possible.


---

### FAQ

**Q: Is this a production implementation?**
A: The README frames it as a teaching repo; use it to learn patterns, then harden the pieces you need.

**Q: What should I run first?**
A: Start with s01 to see a minimal loop, then progress to the later stages for worktrees, teams, and persistence.

**Q: How do I keep it safe?**
A: Use a dedicated API key with quotas, and run demos in a sandbox repo or throwaway project.

---

## Source & Thanks

> GitHub: https://github.com/shareAI-lab/learn-claude-code
> Owner avatar: https://avatars.githubusercontent.com/u/189210346?v=4
> License (SPDX): MIT
> Stars (verified via `api.github.com/repos/shareAI-lab/learn-claude-code`): 59,755


---

<!-- ZH -->

# learn-claude-code——从零搭建 Agent Harness

> learn-claude-code 是 0→1 的 agent harness 教学项目：工具调度、worktree 任务隔离、上下文压缩与团队协议。clone 后配置 ANTHROPIC_API_KEY，运行示例脚本即可逐章复现与调试。

## 快速使用

1. 安装（以仓库为准）：

```bash
git clone https://github.com/shareAI-lab/learn-claude-code && cd learn-claude-code && pip install -r requirements.txt
```

2. 首次运行/检查：

```bash
cp .env.example .env && python agents/s01_agent_loop.py
```

3. 可重复验证：

```bash
python agents/s12_worktree_task_isolation.py
```

---

## 简介

learn-claude-code 已核验来源于 `shareAI-lab/learn-claude-code`（59,755 stars，许可证 MIT）。适合：在做 agent 产品、需要“可运行的参考实现”而不是泛泛文章的工程师。兼容：Python + pip；可选 web UI（npm）用于可视化与文档浏览。装机时间：12 分钟。

### 量化信息

- 多阶段可运行脚本（s01→s12，仓库）
- 装机约 12 分钟

---

## 深度说明

### 解决什么问题

把“找到项目 → 安装 → 首次验证 → 回滚/升级”变成可复用流程，避免只靠一次性提示词或截图操作。

### 最小心智模型

- GitHub 仓库是事实来源：安装、配置、升级、已知问题都以它为准。
- 第一次只做最小闭环：一个命令安装、一个命令验证、一个可回滚方案。
- 记录基线：装机耗时、首次成功输出、完成 1 个真实任务的证据。

### 安全推广清单

1. **核验来源**：确认仓库 URL / stars / license 与预期一致。
2. 按“快速使用”完成安装与首次检查。
3. 跑通“验证命令”，把输出留档。
4. 标注 owner，并写清升级/回滚步骤。

### 常见排障

- **装了但看不到效果**
  - 常见原因：需要重启 CLI/IDE 或配置文件路径不对。
  - 处理：重启后再跑一遍“验证命令”。

- **一台机器能跑，另一台不行**
  - 常见原因：Node/Python/Docker 版本差异或缺少系统依赖。
  - 处理：固定版本，并拷贝最小可用配置。

- **token 成本/延迟超预期**
  - 常见原因：工具 schema 或输出过长被塞进上下文。
  - 处理：拆小步骤、缓存结果、尽量让工具输出更短。


---

### FAQ

**Q: 这是生产级实现吗？**
A: README 明确它是教学仓库；用来理解机制，再把你需要的部分做权限、观测和错误处理加固。

**Q: 先跑哪个最合适？**
A: 从 s01 看最小 loop，再逐步跑到后面的 worktree、teams 与 persistence 章节。

**Q: 怎么更安全地实验？**
A: 用独立额度的 API key，并在 sandbox/一次性项目里跑，避免污染真实仓库。

---

## 来源与感谢

> GitHub：https://github.com/shareAI-lab/learn-claude-code
> Owner avatar：https://avatars.githubusercontent.com/u/189210346?v=4
> 许可证（SPDX）：MIT
> GitHub stars（已通过 `api.github.com/repos/shareAI-lab/learn-claude-code` 核验）：59,755


---
Source: https://tokrepo.com/en/workflows/learn-claude-code-build-an-agent-harness
Author: Agent Toolkit