This repository contains the source code, examples, and project implementations for the book — a practical, code-first guide to building AI agents and, more importantly, putting them into production.
概览
This repository contains the source code, examples, and project implementations for the book — a practical, code-first guide to building AI agents and, more importantly, putting them into production.
README
Atlas Agents — Hands-on AI Agents in Production
This repository contains the source code, examples, and project implementations for the book “Hands-on AI Agents” — a practical, code-first guide to building AI agents and, more importantly, putting them into production.
🚀 Book Overview
Where Agentic Design Patterns taught how to write agents, this book is about how to run them: the harnesses, guardrails, evals, loops, and deployment machinery that turn a demo into a system you can trust unattended. One project — Atlas, an autonomous engineering assistant — grows chapter by chapter from a fifty-line ReAct loop into a self-correcting, self-improving production system.
Key Technologies
- Claude API / OpenAI / Gemini: multi-provider agent cores with structured outputs.
- LangGraph & CrewAI: stateful agent graphs and role-based multi-agent orchestration.
- MCP & A2A: universal tool connectivity and agent-to-agent discovery.
- Claude Code & Antigravity: agentic coding harnesses and loop primitives.
- Agent Skills: declarative, progressively-disclosed expertise (
SKILL.md). - E2B / Docker sandboxes: safe code execution boundaries.
- Managed Agents: server-run sessions, outcomes, and scheduled deployments.
- LiteLLM, DSPy, Ollama: model portability, routing, and local inference.
📁 Repository Structure
Each chapter folder holds the chapter’s Atlas project. Extended examples that go beyond the printed text live in each chapter’s online/ subfolder.
| Folder | Chapter |
|---|---|
ch01_react_from_scratch/ |
Anatomy of an Agent — the minimal ReAct loop |
ch02_prompt_architecture/ |
Prompt Architecture for Agents |
ch03_tools_and_skills/ |
Tools, Skills, and Structured Outputs |
ch04_handoffs/ |
Handoffs and Routines — the support triage router |
ch05_state_graphs/ |
Stateful Agent Graphs — LangGraph persistence and HITL |
ch06_multi_agent/ |
Multi-Agent Collaboration — CrewAI and debate protocols |
ch07_model_portability/ |
One Agent, Many Models — LiteLLM, Ollama, DSPy |
ch08_mcp_a2a/ |
Open Protocols — MCP servers and A2A discovery |
ch09_agent_skills/ |
Agent Skills — the production skill library |
ch10_claude_code_antigravity/ |
Claude Code and Antigravity |
ch11_memory/ |
Memory and Agentic RAG |
ch12_sandboxes/ |
Code Execution and Sandbox Agents |
ch13_multimodal/ |
Multimodal and Voice Agents |
ch14_guardrails/ |
Guardrails and Agent Safety |
ch15_agent_harness/ |
Agent Harness Engineering |
ch16_always_on_agents/ |
Always-On Agents — daemons, watchdogs, recovery |
ch17_managed_agents/ |
Managed Agents — let the platform run it |
ch18_evaluation/ |
Evaluation and Observability |
ch19_deployment/ |
Deployment, Async Agents, and Security |
ch20_loop_engineering/ |
Loop Engineering — the self-correcting fix loop |
ch21_harness_engineer/ |
The Harness Engineer — /learn, adversarial pairs, prose verifiers |
ch22_capstone/ |
Capstone: Atlas — the Autonomous Engineering Assistant |
ch23_future/ |
What’s Next — scaffold optimization and reasoning benchmarks |
shared/ |
Global config and declarative skill models used across chapters |
🛠️ Prerequisites
- Python 3.10+
- API keys as needed per chapter: Anthropic, OpenAI, Google Gemini (see
shared/config.py— keys load from a.envat the repo root) pip install -r requirements.txt(per-chapter extras are noted in each file’s header)- Basic understanding of LLM prompting and Python
▶️ Running the Examples
Every script is self-contained and documents its own usage and dependencies in its module docstring:
cd ch20_loop_engineering
python fix_loop.py --repo ./orders-service --goal "pytest green, ruff clean"
推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add agulli/atlas-agents