Open SWE Framework Released

LangChain today released Open SWE, an open-source framework designed to help engineering organizations build internal coding agents. The framework captures architectural patterns that have emerged across production deployments at Stripe, Ramp, and Coinbase—systems built independently but converging on similar technical solutions.

Key Architectural Patterns

Open SWE implements five core patterns observed in production coding agent systems:

• Isolated execution environments: Tasks run in dedicated cloud sandboxes with full shell access and strict boundaries, containing the blast radius of any errors
• Curated toolsets: Rather than accumulating hundreds of tools, the framework provides a focused set: execute, fetch_url, http_request, commit_and_open_pr, linear_comment, and slack_thread_reply
• Slack-first invocation: Integrates directly into existing communication workflows rather than requiring new interfaces
• Rich startup context: Pulls full context from Linear issues, Slack threads, or GitHub PRs upfront, reducing overhead from exploratory tool calls
• Subagent orchestration: Decomposes complex tasks using child agents with isolated context and focused responsibilities

Technical Implementation

Open SWE is built on LangChain's Deep Agents framework and LangGraph, providing several features:

• Multiple sandbox backends: Supports Modal, Daytona, Runloop, and LangSmith, with extensibility for custom implementations
• Context engineering: Reads optional AGENTS.md files from repositories to encode conventions and team-specific patterns alongside task-specific context from tickets or threads
• Middleware orchestration: Deterministic middleware hooks allow injecting follow-up messages and controlling agent behavior mid-execution
• Composition over forking: Built as a layer on Deep Agents, allowing organizations to adopt framework improvements without rebuilding customizations

Getting Started

Organizations exploring internal coding agents can use Open SWE as a starting point. The framework is customizable for team-specific tools, prompts, and workflows without requiring modifications to core agent logic. The GitHub repository provides implementation details and examples.