LangChain has released Open-SWE, an open-source framework for building internal coding agents similar to those deployed by elite engineering organizations like Stripe, Ramp, and Coinbase. The framework provides the same architectural patterns these companies built internally: cloud sandboxes, Slack and Linear integration, and automatic pull request creation.
"Elite engineering orgs like Stripe, Ramp, and Coinbase are building their own internal coding agents—Slackbots, CLIs, and web apps that meet engineers where they already work," LangChain explains. "Open-SWE is the open-source version of this pattern."
Built on LangChain's LangGraph and Deep Agents frameworks, Open-SWE gives developers the same architecture those companies built internally, but without starting from scratch. It handles cloud sandboxes, Slack and Linear invocation, subagent orchestration, and automatic PR creation—ready to customize for any codebase.
At its core, Open-SWE composes on the Deep Agents framework rather than forking an existing agent. This provides an upgrade path where developers can pull in upstream improvements while customizing orchestration, tools, and middleware for their organization. The approach mirrors how Ramp built on top of OpenCode.
Every task runs in its own isolated cloud sandbox—a remote Linux environment with full shell access. The repository gets cloned in, the agent receives full permissions, and the blast radius of any mistake stays fully contained. No production access, no confirmation prompts.
Open-SWE supports multiple sandbox providers out of the box: Modal, Daytona, Runloop, and LangSmith. Developers can plug in their own providers as well. Each thread gets a persistent sandbox that's reused across follow-up messages, and sandboxes automatically recreate if they become unreachable. Multiple tasks run in parallel, each in its own sandbox with no queuing.
Following Stripe's principle that tool curation matters more than tool quantity, Open-SWE ships with a small, focused toolset: shell commands via execute, web page fetching with fetchurl, API calls through httprequest, and git operations with commitandopen_pr. It also includes Linear and Slack integration tools, plus the built-in Deep Agents tools for file operations.
Context engineering happens through two sources. If the repository contains an AGENTS.md file at the root, it's read from the sandbox and injected into the system prompt—equivalent to Stripe's rule files for encoding conventions, testing requirements, and architectural decisions. The framework also assembles full Linear issue details or Slack thread history for rich initial context.
Orchestration works on two layers. The Deep Agents framework natively supports spawning child agents via the task tool, letting the main agent fan out independent subtasks to isolated subagents. Deterministic middleware hooks run around the agent loop, including checkmessagequeuebeforemodel for injecting follow-up messages and openprif_needed as a safety net that commits and opens PRs if the agent didn't.
Invocation is straightforward: mention the bot in Slack threads with optional repo:owner/name syntax, comment @openswe on Linear issues, or tag @openswe in GitHub PR comments for review feedback. Each invocation creates a deterministic thread ID, so follow-up messages on the same issue or thread route to the same running agent.
The agent is instructed to run linters, formatters, and tests before committing, with the openprif_needed middleware acting as a backstop. If the agent finishes without opening a pull request, the middleware handles it automatically.
Open-SWE includes several standout features that make it production-ready. It reacts with a 👀 emoji the moment it picks up a message, provides instant acknowledgement. Developers can send follow-up messages mid-task and the agent will incorporate them before its next step. Multiple tasks run in parallel without queuing, and GitHub OAuth is built-in for seamless authentication.
Perhaps most importantly, the framework automatically opens draft pull requests when tasks complete, linking back to the originating Linear ticket or Slack thread. This creates a natural review workflow where human developers can inspect changes before merging.
The framework is fully customizable. Organizations can swap the sandbox provider, change models, add tools, modify triggers, adjust the system prompt, and add middleware hooks. The customization guide provides detailed instructions for tailoring Open-SWE to specific organizational needs.
By open-sourcing these patterns, LangChain is democratizing the internal tooling that previously only elite engineering organizations could build. Any development team can now deploy a coding agent that operates in isolated sandboxes, integrates with existing workflows, and handles the tedious parts of software development—all while maintaining appropriate safety boundaries.
Questions and Answers
What is Open-SWE?
Open-SWE is an open-source framework from LangChain for building internal coding agents. It provides the same architectural patterns used by companies like Stripe, Ramp, and Coinbase: cloud sandboxes for isolated execution, integration with Slack and Linear, subagent orchestration, and automatic GitHub pull request creation. It's built on LangGraph and Deep Agents frameworks.
How do isolated sandboxes work in Open-SWE?
Every task runs in its own remote Linux environment with full shell access. The repository is cloned into the sandbox, the agent gets complete permissions, and any mistakes are fully contained. Open-SWE supports Modal, Daytona, Runloop, and LangSmith as sandbox providers. Each thread gets a persistent sandbox that's reused for follow-up messages, and multiple tasks run in parallel without queuing.
How do you invoke Open-SWE agents?
You can trigger Open-SWE in three ways: mention the bot in Slack threads (with optional repo:owner/name syntax to specify which repository), comment @openswe on Linear issues, or tag @openswe in GitHub PR comments. The agent reacts with 👀 to acknowledge, reads full context from the issue or thread, and posts results back with links to draft pull requests.
What makes Open-SWE different from other coding agents?
Open-SWE follows the architectural decisions of elite internal tools rather than consumer-focused coding assistants. It prioritizes isolation (sandboxed execution), curation (small focused toolset), and workflow integration (Slack, Linear, GitHub). The framework is fully customizable—you can swap models, tools, middleware, and sandbox providers to match your organization's needs while keeping the proven architectural patterns.
