LangChain Releases Open SWE Framework for Enterprise AI Coding Agents
Rongchai Wang Mar 17, 2026 17:33
LangChain open-sources Open SWE, a framework mirroring coding agent architectures deployed at Stripe, Coinbase, and Ramp. Built on Deep Agents and LangGraph.
LangChain has released Open SWE, an open-source framework capturing the architectural patterns that Stripe, Coinbase, and Ramp independently developed for their internal AI coding agents. The MIT-licensed project, built on LangChain's Deep Agents and LangGraph platforms, provides a customizable foundation for engineering organizations looking to deploy autonomous coding assistants.
Enterprise Convergence Drives the Design
The framework emerges from observable convergence among major fintech players. Stripe built Minions, Ramp developed Inspect, and Coinbase created Cloudbot—each arriving at similar architectural decisions despite working independently.
Those shared patterns include isolated cloud sandboxes for code execution, curated toolsets (Stripe reportedly maintains around 500 carefully selected tools), Slack-first invocation, rich context injection from Linear issues or GitHub PRs, and subagent orchestration for complex tasks.
"These architectural choices have proven effective across multiple production deployments," LangChain noted in the announcement, though they acknowledge organizations will need to adapt components to their own environments.
Technical Architecture
Open SWE ships with approximately 15 curated tools covering shell execution, web fetching, API calls, Git operations, and integrations with Linear and Slack. The framework supports pluggable sandbox providers including Modal, Daytona, Runloop, and LangSmith.
Each task runs in an isolated Linux environment with full shell access. The repository gets cloned in, the agent receives complete permissions within that boundary, and errors remain contained. Multiple tasks can run in parallel, each in separate sandboxes.
Context engineering happens through two channels: an AGENTS.md file at the repository root encoding team conventions and architectural decisions, plus full Linear issue or Slack thread history assembled before the agent begins work.
The orchestration layer combines model-driven subagent spawning with deterministic middleware hooks. One middleware component injects follow-up messages that arrive mid-run. Another acts as a safety net, automatically committing and opening a PR if the agent doesn't complete that step.
Composition Over Forking
Rather than forking an existing agent, Open SWE composes on the Deep Agents framework—similar to how Ramp's team built Inspect on top of OpenCode. This approach provides an upgrade path: when Deep Agents improves context management or token efficiency, those improvements can flow through without rebuilding customizations.
Deep Agents handles file-based memory to prevent context overflow on larger codebases, provides structured planning via a write_todos tool, and supports isolated subagent spawning where different subtasks don't pollute each other's conversation history.
How It Compares
The comparison to enterprise implementations reveals expected differences in implementation details. Stripe uses forked Goose with AWS EC2 devboxes and three-layer validation. Ramp composed on OpenCode with Modal containers and visual DOM verification. Coinbase built from scratch with agent councils and auto-merge capabilities.
Open SWE defaults to Claude Opus 4 but supports any LLM provider. Organizations can configure different models for different subtasks.
Deployment Reality
The framework represents LangChain's bet on a specific trajectory for AI-assisted development: autonomous, long-running agents that integrate with existing developer workflows rather than requiring new interfaces. This differs from the short, synchronous, in-IDE copilot model that dominated earlier AI coding tools.
Documentation includes an installation guide covering GitHub App creation, LangSmith setup, and production deployment, plus a customization guide for swapping sandbox providers, models, tools, and triggers.
Open SWE is available now at github.com/langchain-ai/open-swe. Organizations interested in the LangSmith Sandboxes can join a waitlist through LangChain's website.
Image source: Shutterstock.jpg)