LangChain 的开源软件工程代理工厂

Brace 2026-07-19 8.6 KB 187.9K 303 𝕏 文章列表

LangChain 推出一套开源软件工程代理工具，包括本地编码的 dcode、云端自动作业的 OpenSWE、代码审查的 OpenSWE Review 以及维护文档的 OpenWiki。这些工具基于 Deep Agents 构建，支持开源模型和 LangSmith 追踪，旨在让团队在不同工作流中拥有对 AI 代理的完全控制与定制能力。

LangChain 开源代码审查 OpenSWE dcode AI代理

# LangChain’s Open-Source Software Factory **作者**: Brace **日期**: 2026-07-18T19:13:42.000Z **来源**: [https://x.com/BraceSproul/status/2078558852921094253](https://x.com/BraceSproul/status/2078558852921094253) --- ![Image](/doc/3154/image/img_c02f0374f350.jpg) At LangChain, we interact with software engineering agents from the terminal, Slack, Linear, GitHub, and CI. Different workflows need different agents. Local coding should feel interactive. Cloud coding should run in the background and open PRs. Code review should be careful and tied into GitHub. Repo documentation should stay close to the code and give agents useful context. We’ve built that system as a set of open-source tools: - dcode for local coding - OpenSWE for cloud coding agents - OpenSWE Review for automated code review - OpenWiki for repo documentation and agent memory Together, they form our software engineering agent factory. This is the shape we use internally: local agents, cloud agents, review agents, repo memory, open models, and LangSmith traces across the system. ## Why open source matters Software engineering agents read code, edit code, review PRs, and participate in engineering workflows. Teams need to understand and control how those agents behave. That is hard if the agent is a closed system. You are limited to the provider’s supported models, review behavior, integrations, and observability. We want agents that teams can inspect, modify, and adapt to their own workflows, including their own review standards, repo conventions, model choices, and internal tools. That is why these projects are open source. ## The pieces dcode dcode is the local coding agent we use from the terminal. ![Image](/doc/3154/image/img_0324f09bda54.jpg) The CLI and agent runtime run locally on your machine, while the model can still be hosted. That lets developers work directly against a local repo while using large coding models that may not run on a laptop. We use dcode for interactive coding work: exploring a repo, making edits, and running implementation tasks from the terminal. OpenSWE OpenSWE is the cloud coding agent we use when work should happen in the background. A lot of engineering work starts outside the terminal. A bug comes up in Slack. A feature request lives in Linear. A repo maintenance task needs to run on a schedule. OpenSWE can be triggered from Slack, Linear, GitHub, or the web UI. It runs in the cloud, works on the code, and can open a pull request when it is done. Internally, this has become one of our most-used agent workflows. In just the last week, we triggered OpenSWE almost 1,000 times from Slack alone, not including Linear or UI usage. Because OpenSWE runs in a sandbox in the cloud, we can run many coding agents in parallel without tying up anyone’s local environment. OpenSWE also includes a UI for inspecting agent work, chatting with the cloud coding agent, configuring models and prompts, viewing analytics, and setting up scheduled automations. ![Image](/doc/3154/image/img_b27caf334c5a.jpg) OpenSWE Review OpenSWE Review connects to GitHub and automatically reviews pull requests. It identifies bugs and regressions before code merges. ![Image](/doc/3154/image/img_2da7d8a41fc1.jpg) It also performs extremely well on external benchmarks. On the Offline Code Review Benchmark, OpenSWE Review scores 47% when running GPT-5.5 with medium reasoning. That places it #6 overall and #1 among open-source code review agents. Code review is highly organization-specific. Teams differ in what issues they care about, how they phrase comments, and when they want an agent to block, suggest, or stay quiet. OpenSWE Review gives us a strong default that we can adapt by changing prompts, workflows, model choices, and repo-specific behavior. OpenWiki OpenWiki generates and maintains codebase documentation using Google’s Open Knowledge Format. The docs live with the code and are kept up to date automatically via a GitHub action. That gives both humans and agents a maintained knowledge layer for the repo, without needing to worry about maintaining the docs (OpenWiki does that for you!). ![Image](/doc/3154/image/img_66760c57762f.jpg) If an agent has to rediscover the architecture of a repo every time it runs, it wastes tokens and misses important conventions. OpenWiki gives coding and review agents a better starting point, while supporting an auto-update flow so you never need to think about maintaining your agent docs again. ## The stack Deep Agents dcode, OpenSWE, OpenSWE Review, and OpenWiki are all built on Deep Agents. The workflows are different, but they share the same underlying agent foundation. That gives us a consistent way to build, customize, and improve agents across local coding, cloud coding, review, and documentation. Sharing code, resources and engineering know-how is made even easier when they’re all built in the same framework. Open models We’ve optimized the stack for open models. SWE agents get expensive quickly when they run across an organization. Local coding, cloud coding, code review, documentation, scheduled maintenance, and repeated repo analysis all use tokens. Open models give us more control over cost (the key reason), latency, and deployment choices. They also let teams route different tasks to different models, experiment with open coding models, or even fine-tune models for their own repos and workflows. LangSmith We trace these agents with LangSmith. When an agent opens a PR, leaves a review comment, or fails on a task, traces show the files it inspected, the context it loaded, the tool calls it made, and where it got stuck. Those traces help us debug individual runs and improve the system over time. We use them to identify failure modes, improve prompts, compare models, and feed higher-quality examples into continual improvement workflows with LangSmith Engine. Engine for coding agents is a new workflow we’ve been experimenting with, where we have Engine run over coding agent traces, identify shortfalls, and propose optimizations. Since every coding agent run from every employee at LangChain is traced, Engine is able to do identify and preform these optimizations from birds eye view, instead of individually on specific engineers traces. ## Build your own software engineering agent factory This stack reflects how we use software engineering agents at LangChain, but the pieces are open source so other teams can adapt them, and retain full control over your own software factory. You can start small: 1. Use dcode for local coding. 2. Add OpenSWE to run coding agents from Slack, Linear, GitHub, or the web UI. 3. Turn on OpenSWE Review for automated PR review. 4. Use OpenWiki to maintain repo knowledge for humans and agents. 5. Connect traces to LangSmith if you want observability and improvement loops across the system. The goal is not to move every engineering workflow into one agent interface. It is to put the right agents in the places where engineering work already happens, with enough control to make them fit your repos, models, and team conventions. ## 相关链接 - [Brace](https://x.com/BraceSproul) - [@BraceSproul](https://x.com/BraceSproul) - [99K](https://x.com/BraceSproul/status/2078558852921094253/analytics) - [dcode](https://docs.langchain.com/oss/python/deepagents/code/overview) - [OpenSWE](https://github.com/langchain-ai/open-swe) - [OpenSWE Review](https://github.com/langchain-ai/open-swe/blob/main/agent/reviewer.py) - [OpenWiki](https://github.com/langchain-ai/openwiki) - [dcode](https://docs.langchain.com/oss/python/deepagents/code/overview) - [OpenSWE](https://github.com/langchain-ai/open-swe) - [Offline Code Review Benchmark](https://codereview.withmartian.com/?mode=offline) - [OpenWiki](https://github.com/langchain-ai/openwiki) - [Open Knowledge Format](https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md) - [LangSmith](https://langsmith.com/) - [LangSmith Engine](https://www.langchain.com/langsmith/engine) - [dcode](https://docs.langchain.com/oss/python/deepagents/code/overview) - [OpenSWE](https://github.com/langchain-ai/open-swe) - [OpenWiki](https://github.com/langchain-ai/openwiki) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [3:13 AM · Jul 19, 2026](https://x.com/BraceSproul/status/2078558852921094253) - [99.1K Views](https://x.com/BraceSproul/status/2078558852921094253/analytics) - [View quotes](https://x.com/BraceSproul/status/2078558852921094253/quotes) --- *导出时间: 2026/7/19 14:43:33* --- ## 中文翻译 # LangChain 的开源软件工厂 **作者**: Brace **日期**: 2026-07-18T19:13:42.000Z **来源**: [https://x.com/BraceSproul/status/2078558852921094253](https://x.com/BraceSproul/status/2078558852921094253) --- ![Image](/doc/3154/image/img_c02f0374f350.jpg) 在 LangChain，我们通过终端、Slack、Linear、GitHub 和 CI 与软件工程智能体交互。不同的工作流需要不同的智能体。本地编码应该具有交互性。云端编码应该在后台运行并创建 PR。代码审查应该严谨并与 GitHub 紧密集成。仓库文档应该紧贴代码，并为智能体提供有用的上下文。我们将该系统构建为一套开源工具： - dcode：用于本地编码 - OpenSWE：用于云端编码智能体 - OpenSWE Review：用于自动代码审查 - OpenWiki：用于仓库文档和智能体记忆它们共同构成了我们的软件工程智能体工厂。这是我们内部使用的形态：本地智能体、云端智能体、审查智能体、仓库记忆、开放模型，以及贯穿整个系统的 LangSmith 追踪。 ## 为什么开源很重要软件工程智能体读取代码、编辑代码、审查 PR 并参与工程工作流。团队需要理解并控制这些智能体的行为。如果智能体是一个封闭系统，这很难做到。你受限于提供商支持的模型、审查行为、集成和可观测性。我们希望团队可以检查、修改并调整智能体以适应自己的工作流，包括他们自己的审查标准、仓库约定、模型选择和内部工具。这就是这些项目开源的原因。 ## 组成部分 dcode dcode 是我们在终端中使用的本地编码智能体。 ![Image](/doc/3154/image/img_0324f09bda54.jpg) CLI（命令行界面）和智能体运行时在您的本地机器上运行，而模型仍然可以托管在云端。这让开发者能够直接针对本地仓库工作，同时使用那些可能无法在笔记本电脑上运行的大型编码模型。我们使用 dcode 进行交互式编码工作：探索仓库、进行编辑，以及从终端执行实现任务。 OpenSWE OpenSWE 是我们在工作应该在后台进行时使用的云端编码智能体。许多工程工作开始于终端之外。Slack 上出现了一个错误报告。Linear 中存在一个功能请求。一项仓库维护任务需要按计划运行。 OpenSWE 可以由 Slack、Linear、GitHub 或 Web UI 触发。它在云端运行，处理代码，并在完成后可以打开一个拉取请求（PR）。在内部，这已成为我们最常用的智能体工作流之一。仅在上周，我们就仅从 Slack 触发了 OpenSWE 近 1,000 次，还不包括在 Linear 或 UI 中的使用。因为 OpenSWE 在云端的沙盒中运行，我们可以并行运行许多编码智能体，而不会占用任何人的本地环境。 OpenSWE 还包含一个用于检查智能体工作、与云端编码智能体聊天、配置模型和提示词、查看分析以及设置计划自动化的 UI。 ![Image](/doc/3154/image/img_b27caf334c5a.jpg) OpenSWE Review OpenSWE Review 连接到 GitHub 并自动审查拉取请求。它在代码合并之前识别错误和回归。 ![Image](/doc/3154/image/img_2da7d8a41fc1.jpg) 它在外部基准测试中也表现极其出色。在 Offline Code Review Benchmark（离线代码审查基准）上，OpenSWE Review 在运行具有中等推理能力的 GPT-5.5 时得分 47%。这使其在总排名中位列第 6，在开源代码审查智能体中位列第 1。代码审查具有高度的组织特异性。团队在关注哪些问题、如何撰写评论，以及希望智能体何时阻止、建议或保持沉默方面各不相同。 OpenSWE Review 为我们提供了一个强大的默认配置，我们可以通过更改提示词、工作流、模型选择和特定于仓库的行为来调整它。 OpenWiki OpenWiki 使用 Google 的 Open Knowledge Format（开放知识格式）生成和维护代码库文档。文档与代码共存，并通过 GitHub action 自动保持更新。这为人类和智能体都提供了一个维护良好的仓库知识层，而无需担心维护文档的问题（OpenWiki 会为您代劳！）。 ![Image](/doc/3154/image/img_66760c57762f.jpg) 如果智能体每次运行时都必须重新发现仓库的架构，它会浪费 Token 并错过重要的约定。OpenWiki 为编码和审查智能体提供了一个更好的起点，同时支持自动更新流程，因此您无需再考虑维护您的智能体文档。 ## 技术栈 Deep Agents（深度智能体） dcode、OpenSWE、OpenSWE Review 和 OpenWiki 都是基于 Deep Agents 构建的。工作流不同，但它们共享相同的基础智能体框架。这为我们提供了一种在本地编码、云端编码、审查和文档之间构建、定制和改进智能体的一致方法。当它们都构建在同一个框架中时，共享代码、资源和工程知识变得更加容易。 Open models（开放模型）我们针对开放模型优化了技术栈。当 SWE（软件工程）智能体在整个组织中运行时，成本会迅速上升。本地编码、云端编码、代码审查、文档、计划维护和重复的仓库分析都消耗 Token。开放模型让我们能更好地控制成本（主要原因）、延迟和部署选择。它们还允许团队将不同的任务路由到不同的模型，尝试开放的编码模型，甚至针对自己的仓库和工作流微调模型。 LangSmith 我们使用 LangSmith 追踪这些智能体。当智能体打开 PR、留下审查评论或在任务上失败时，追踪会显示它检查的文件、加载的上下文、进行的工具调用以及它卡住的地方。这些追踪帮助我们调试单个运行并随时间改进系统。我们使用它们来识别故障模式、改进提示词、比较模型，并使用 LangSmith Engine 将更高质量的示例输入到持续改进的工作流中。用于编码智能体的 Engine 是我们一直在试验的一个新工作流，我们让 Engine 运行在编码智能体的追踪之上，识别缺陷并提出优化建议。由于 LangChain 每位员工的每一次编码智能体运行都被追踪，Engine 能够从全局视角（鸟瞰图）识别并执行这些优化，而不是单独针对特定工程师的追踪进行。 ## 构建您自己的软件工程智能体工厂这个技术栈反映了我们在 LangChain 使用软件工程智能体的方式，但这些组成部分是开源的，因此其他团队可以调整它们，并保留对您自己的软件工厂的完全控制。您可以从小处着手： 1. 使用 dcode 进行本地编码。 2. 添加 OpenSWE 以从 Slack、Linear、GitHub 或 Web UI 运行编码智能体。 3. 开启 OpenSWE Review 进行自动 PR 审查。 4. 使用 OpenWiki 为人类和智能体维护仓库知识。 5. 如果您希望在整个系统中具有可观测性和改进循环，请将追踪连接到 LangSmith。目标不是将每个工程工作流都转移到一个智能体界面中。而是将正确的智能体放置在工程工作已经发生的地方，并提供足够的控制，使其适应您的仓库、模型和团队约定。 ## 相关链接 - [Brace](https://x.com/BraceSproul) - [@BraceSproul](https://x.com/BraceSproul) - [99K](https://x.com/BraceSproul/status/2078558852921094253/analytics) - [dcode](https://docs.langchain.com/oss/python/deepagents/code/overview) - [OpenSWE](https://github.com/langchain-ai/open-swe) - [OpenSWE Review](https://github.com/langchain-ai/open-swe/blob/main/agent/reviewer.py) - [OpenWiki](https://github.com/langchain-ai/openwiki) - [dcode](https://docs.langchain.com/oss/python/deepagents/code/overview) - [OpenSWE](https://github.com/langchain-ai/open-swe) - [Offline Code Review Benchmark](https://codereview.withmartian.com/?mode=offline) - [OpenWiki](https://github.com/langchain-ai/openwiki) - [Open Knowledge Format](https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md) - [LangSmith](https://langsmith.com/) - [LangSmith Engine](https://www.langchain.com/langsmith/engine) - [dcode](https://docs.langchain.com/oss/python/deepagents/code/overview) - [OpenSWE](https://github.com/langchain-ai/open-swe) - [OpenWiki](https://github.com/langchain-ai/openwiki) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [3:13 AM · Jul 19, 2026](https://x.com/BraceSproul/status/2078558852921094253) - [99.1K Views](https://x.com/BraceSproul/status/2078558852921094253/analytics) - [View quotes](https://x.com/BraceSproul/status/2078558852921094253/quotes) --- *导出时间: 2026/7/19 14:43:33*