# Symphony — Orchestrate Autonomous Coding Agent Runs

> Symphony orchestrates isolated autonomous runs so teams manage work instead of supervising coding agents. Includes a reference implementation.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

1. Install / run:
   ```bash
   git clone https://github.com/openai/symphony && cd symphony
   ```
2. Start / smoke test:
   ```bash
   sed -n '1,220p' SPEC.md && sed -n '1,220p' elixir/README.md
   ```
3. Verify:
   - Run one small task with the reference implementation and confirm it produces clear proof-of-work artifacts (CI status, review feedback, etc.) as described in the README.

## Intro

Symphony orchestrates isolated autonomous runs so teams manage work instead of supervising coding agents. Includes a reference implementation.

- **Best for:** Teams using coding agents who want repeatable, isolated “implementation runs” with proof-of-work outputs
- **Works with:** Coding agents + CI/test harnesses; best with harness-engineered repos (see repo docs)
- **Setup time:** 20 minutes

## How to Use It Well

- Setup time ~20 minutes (read SPEC + follow `elixir/README.md` for the reference implementation)
- 2 options: implement the SPEC yourself or use the experimental reference implementation
- GitHub stars + forks (verified): see Source & Thanks

If you’re already using coding agents, the biggest unlock is operationalizing them: isolate runs, make inputs/outputs explicit, and require proof-of-work (tests, checks, or review signals) so humans manage work, not prompt loops.

### FAQ

**Q: What is an “implementation run”?**
A: A single isolated agent execution that takes a work item and returns proof of work (outputs you can verify) instead of an open-ended chat loop.

**Q: Do I have to use Elixir?**
A: No. The repo describes a language-agnostic SPEC and also provides an experimental Elixir reference implementation.

**Q: Is it production-ready?**
A: The README labels it an engineering preview; run it only in trusted environments and validate the harness boundaries first.

## Source & Thanks

> Source: https://github.com/openai/symphony
> License: Apache-2.0
> GitHub stars: 23,343 · forks: 2,208

---

<!-- ZH -->

## 快速使用

1. 安装 / 运行：
   ```bash
   git clone https://github.com/openai/symphony && cd symphony
   ```
2. 启动 / 冒烟测试：
   ```bash
   sed -n '1,220p' SPEC.md && sed -n '1,220p' elixir/README.md
   ```
3. 验证：
   - Run one small task with the reference implementation and confirm it produces clear proof-of-work artifacts (CI status, review feedback, etc.) as described in the README.

## 简介

Symphony 把任务拆成“隔离的自主实现运行”，让团队管理工作而不是盯着 coding agent。它强调 harness engineering 思路，并提供参考实现与 SPEC，适合在可信环境试运行并沉淀可复用运行模板。

- **适合谁：** 希望把“实现任务”变成可重复、可隔离运行并产出证据的团队（而不是手把手盯着 agent）
- **可搭配：** 编码代理 + CI/测试 harness；在采用 harness engineering 的仓库里更合适（见仓库文档）
- **准备时间：** 20 分钟

## 实战建议

- 上手约 20 分钟（读 SPEC + 按 `elixir/README.md` 跑参考实现）
- 两条路径：自己按 SPEC 实现，或使用实验性的参考实现
- GitHub stars + forks（已核验）：见「来源与感谢」

如果你已经在用 coding agent，关键不是“更会提示词”，而是工程化：隔离运行、明确输入输出，并要求可验证的 proof-of-work（测试/检查/审阅信号），让人类管理工作而不是管理对话。

### FAQ

**什么是“implementation run”？**
A: 一次隔离的 agent 执行：输入一个工作项，输出可验证的结果与证据，而不是无限对话循环。

**一定要用 Elixir 吗？**
A: 不需要。仓库提供语言无关的 SPEC，同时给了实验性的 Elixir 参考实现。

**能直接上生产吗？**
A: README 明确是工程预览版；建议只在可信环境试跑，并先把 harness 边界与权限收紧。

## 来源与感谢

> Source: https://github.com/openai/symphony
> License: Apache-2.0
> GitHub stars: 23,343 · forks: 2,208


---
Source: https://tokrepo.com/en/workflows/symphony-orchestrate-autonomous-coding-agent-runs
Author: Agent Toolkit