<p>LangChain released <a href="https://github.com/langchain-ai/open-swe" target="_blank" rel="noopener nofollow">Open SWE</a>, an MIT-licensed framework that codifies the AI coding agent architecture that Stripe (Minions), Ramp (Inspect), and Coinbase (Cloudbot) each built independently. The idea is straightforward: these companies spent real engineering time figuring out how to run coding agents in production, and those patterns are worth packaging up so everyone else doesn’t have to start from scratch.</p>
<h2>How it’s built</h2>
<p>Open SWE runs tasks in isolated cloud sandboxes with full shell access. It ships with around 15 curated tools, including web search on top of the standard set from LangChain’s Deep Agents. Context is managed through an <code>AGENTS.md</code> file injected into the system prompt - a simple but effective way to give the agent project-specific knowledge without baking it into the framework itself.</p>
<p>For orchestration, it supports subagents to break complex tasks into smaller pieces. Validation happens inside the agent rather than as a separate layer.</p>
<p>The sandbox setup is pluggable - you can bring your own or use LangSmith sandboxes, with integrations for <a href="https://e2b.dev" target="_blank" rel="noopener nofollow">E2B</a>, <a href="https://daytona.io" target="_blank" rel="noopener nofollow">Daytona</a>, <a href="https://modal.com" target="_blank" rel="noopener nofollow">Modal</a>, and <a href="https://runloop.ai" target="_blank" rel="noopener nofollow">Runloop</a>. It defaults to Claude Opus 4.6 but works with any LLM provider.</p>
<h2>Middleware for the boring-but-important stuff</h2>
<p>One of the more practical design decisions is the middleware architecture. Deterministic CI gates, security scans, and approval workflows can be added without touching core agent logic. This matters in practice - the things that make agents safe to run in a real company (audit trails, human-in-the-loop approvals, security checks) shouldn’t require forking the agent itself.</p>
<h2>Invocation</h2>
<p>Agents can be triggered from Slack, Linear, or GitHub, which covers most of where engineering work actually gets tracked.</p>
<p>The framework is built on LangChain’s Deep Agents and LangGraph. The full comparison of how it stacks up against the Stripe/Ramp/Coinbase implementations is broken down in <a href="https://twitter.com/kishan_dahya" target="_blank" rel="noopener nofollow">this analysis by Kishan Dahya</a>.</p>


Collections

LangChain released Open SWE, an MIT-licensed open-source framework for building internal coding agents, inspired by production implementations at Stripe (Minions), Ramp (Inspect), and Coinbase (Cloudbot). It runs tasks in isolated cloud sandboxes with shell access, ships with ~15 curated tools, and uses an AGENTS.md file for project-specific context injection. The framework supports subagents for complex task decomposition, a pluggable sandbox setup (E2B, Daytona, Modal, Runloop), and defaults to Claude Opus 4.6 while supporting any LLM. A middleware architecture handles CI gates, security scans, and approval workflows without touching core agent logic. Agents can be triggered from Slack, Linear, or GitHub, and the framework is built on LangChain's Deep Agents and LangGraph.

Open SWE: LangChain's open-source framework for internal coding agents