🛠 Open SWE：LangChain 出品 9.8k Stars 的开源内部编码 Agent 框架，Stripe/Ramp/Coinbase 同款架构

🛠 Open SWE：LangChain 出品 9.8k Stars 的开源内部编码 Agent 框架，Stripe/Ramp/Coinbase 同款架构

Stripe 的 Minions、Ramp 的 Inspect、Coinbase 的 Cloudbot——这些顶级科技公司都在搞自己的内部编码 Agent，跑在 Slack 里、绑着 Linear、自动提 PR。现在 LangChain 把同样一套架构开源了：Open SWE，基于 LangGraph + Deep Agents，一周内冲到 9.8k Stars。

不是什么？

不是又一个 Claude Code 替代品。Open SWE 是一套框架，让你给自己的团队搭一个专属的编码 Agent——连上你们自己的 repo、Slack 频道、Linear 看板，agent 用完了自动提 PR，人在 Slack 里回一句就接着干。

核心架构

Open SWE 的架构跟 Stripe/Ramp/Coinbase 保持一致：

1. Agent Harness — 基于 Deep Agents 组合（不是 fork），上游更新能直接 pull，同时保留你的定制空间：

create_deep_agent(
    model="openai:gpt-5.5",
    system_prompt=construct_system_prompt(...),
    tools=[http_request, fetch_url, linear_comment, slack_thread_reply],
    backend=sandbox_backend,
    middleware=[ToolErrorMiddleware(), check_message_queue_before_model, ...],
)

2. 沙箱隔离 — 每次任务跑在独立的云沙箱里（支持 Modal、Daytona、Runloop、LangSmith），agent 有完整 shell 权限但隔离了所有生产环境。

3. 精简化工具集 — Stripe 说得好：工具的质量比数量重要。Open SWE 只给了 15 个左右的工具：execute（跑命令）、fetch_url、http_request、linear_comment、slack_thread_reply，配合 Deep Agents 自带的 read_file/write_file/edit_file 等。

4. 上下文工程 — 如果 repo 根目录有 AGENTS.md，自动注入 system prompt。Linear 的完整 issue 上下文或 Slack 线程历史也一并传入，agent 启动就有全局视野。

5. 编排层 — 主 agent 可以派生子 agent（task 工具），各干各的互不干扰。中间件钩子接管了整个 agent 循环：中途有人发消息（check_message_queue_before_model）、超时通知（notify_step_limit_reached）、工具错误处理（ToolErrorMiddleware）。

怎么触发

快速上手

git clone https://github.com/langchain-ai/open-swe.git
cd open-swe
uv venv
source .venv/bin/activate
uv sync --all-extras

然后注册一个 GitHub App、连上 LangSmith、配好 Slack/Linear webhook，就能用了。详细的安装步骤看 INSTALLATION.md。

一句话总结

如果你团队超过 5 个工程师、每天有修不完的 issue 和 review，Open SWE 可能是最适合你的编码 Agent 框架。不需要从零造轮子——LangChain 把 Stripe/Ramp/Coinbase 踩过的坑都封装好了。

🛠 Open SWE: LangChain's 9.8k Stars Open-Source Internal Coding Agent Framework — Same Architecture as Stripe/Ramp/Coinbase

Stripe has Minions. Ramp has Inspect. Coinbase has Cloudbot. Every elite engineering org is building internal coding agents — Slackbots connected to Linear that auto-open PRs. Now LangChain open-sourced the same pattern: Open SWE, built on LangGraph + Deep Agents, hitting 9.8k stars in its first week.

What it's not

Not another Claude Code fork. Open SWE is a framework for building your team's own coding agent — wired to your repos, Slack channels, and Linear boards. The agent fixes bugs, implements features, opens PRs, and you just reply in Slack to steer it.

Architecture

1. Agent Harness — Composed on Deep Agents (not forked), so you can pull upstream updates while keeping your customizations:

create_deep_agent(
    model="openai:gpt-5.5",
    system_prompt=construct_system_prompt(...),
    tools=[http_request, fetch_url, linear_comment, slack_thread_reply],
    backend=sandbox_backend,
    middleware=[ToolErrorMiddleware(), check_message_queue_before_model, ...],
)

2. Sandbox Isolation — Every task runs in its own cloud sandbox (Modal, Daytona, Runloop, or LangSmith). Full shell access inside, zero blast radius outside.

3. Curated Tools — Stripe's insight: tool curation > tool quantity. ~15 tools including execute, fetch_url, http_request, linear_comment, slack_thread_reply, plus Deep Agents built-ins.

4. Context Engineering — AGENTS.md at repo root gets injected into system prompt automatically. Full Linear issue context or Slack thread history feeds in so the agent starts informed.

5. Middleware & Subagents — Main agent spawns children via the task tool. Middleware hooks handle mid-run messages, timeouts, and tool errors without interrupting the loop.

Invocation

Quick Start

git clone https://github.com/langchain-ai/open-swe.git
cd open-swe
uv venv
source .venv/bin/activate
uv sync --all-extras

Then create a GitHub App, configure LangSmith, and set up Slack/Linear webhooks. Full guide: INSTALLATION.md.

TL;DR

If your team has 5+ engineers and a never-ending stream of issues and PRs, Open SWE is probably the coding agent framework you've been waiting for. No need to reinvent the wheel — LangChain packaged all the lessons from Stripe, Ramp, and Coinbase.