# AG2 — Open-Source AgentOS for Multi-Agent Systems

> AG2 (formerly AutoGen) is an open-source framework for building cooperating AI agents with tool use, human-in-the-loop workflows, and patterns.

## Install

Save as a script file and run:

## Quick Use

1. Install / run:
   ```bash
   pip install ag2[openai]
   ```
2. Start / smoke test:
   ```bash
   python - <<'PY'
import autogen
print('autogen module:', autogen.__name__)
PY
   ```
3. Verify:
   - Run the smallest “first agent” example from the README/docs and confirm you can complete one task end-to-end before adding more agents.

## Intro

AG2 (formerly AutoGen) is an open-source framework for building cooperating AI agents with tool use, human-in-the-loop workflows, and patterns.

- **Best for:** Teams prototyping multi-agent collaboration patterns who want an open-source AgentOS-style framework
- **Works with:** Python workflows; docs and examples are linked from the README
- **Setup time:** 15 minutes

## How to Use It Well

- Setup time ~15 minutes (pip install + run one starter example)
- The project positions itself as AgentOS-style multi-agent tooling (formerly AutoGen)
- GitHub stars + forks (verified): see Source & Thanks

Treat multi-agent as a control problem. Start with a tiny set of roles, define tool boundaries, and log every tool call. Once you can reliably reproduce a run, then scale to more agents and longer horizons.

### FAQ

**Q: Is it stable?**
A: The README discusses roadmap and versioning; start with one pattern and upgrade deliberately.

**Q: How many agents should I start with?**
A: Two is enough: one assistant and one user/tool proxy. Add more only after you can debug failures.

**Q: How do I keep it safe?**
A: Use least-privilege keys, separate dev/prod, and require explicit approvals for any write or shell execution.

## Source & Thanks

> Source: https://github.com/ag2ai/ag2
> License: Apache-2.0
> GitHub stars: 4,533 · forks: 621

---

<!-- ZH -->

## 快速使用

1. 安装 / 运行：
   ```bash
   pip install ag2[openai]
   ```
2. 启动 / 冒烟测试：
   ```bash
   python - <<'PY'
import autogen
print('autogen module:', autogen.__name__)
PY
   ```
3. 验证：
   - Run the smallest “first agent” example from the README/docs and confirm you can complete one task end-to-end before adding more agents.

## 简介

AG2（原 AutoGen）是开源的 AgentOS 框架：提供多 agent 对话、工具调用、HITL（人类在环）与多种对话模式。适合从研究原型起步，逐步把多 Agent 系统收敛成可维护、可扩展的工程项目。

- **适合谁：** 想验证多 Agent 协作模式，并希望从开源 AgentOS 风格框架起步的团队
- **可搭配：** Python 工作流；文档与示例入口见 README
- **准备时间：** 15 分钟

## 实战建议

- 上手约 15 分钟（安装 + 跑通一个最小示例）
- 项目定位为 AgentOS 风格的多 Agent 工具（原 AutoGen）
- GitHub stars + forks（已核验）：见「来源与感谢」

把多 agent 当控制问题：先从很小的角色集合开始，划清工具边界，并记录每次工具调用。能稳定复现一次 run 后，再扩展到更多 agent 与更长任务周期。

### FAQ

**稳定性如何？**
A: README 提到版本路线；建议先用一种模式跑稳，再有计划地升级。

**一开始需要多少个 agent？**
A: 两个就够：一个助手 + 一个用户/工具代理。能排查失败后再加复杂度。

**怎么保证安全？**
A: 用最小权限密钥、区分 dev/prod，并对任何写操作/命令执行加明确审批门。

## 来源与感谢

> Source: https://github.com/ag2ai/ag2
> License: Apache-2.0
> GitHub stars: 4,533 · forks: 621


---
Source: https://tokrepo.com/en/workflows/ag2-open-source-agentos-for-multi-agent-systems
Author: Agent Toolkit