## Introduction Open SWE is an open-source framework by LangChain for building internal coding agents that can autonomously handle software engineering tasks. With 6,500+ stars and an MIT license, it provides the same architectural patterns used by engineering teams at companies like Stripe, Ramp, and Coinbase to build their internal AI coding assistants. The framework enables organizations to trigger automated coding tasks through Slack, Linear, or GitHub, with agents executing in isolated cloud sandboxes and delivering pull requests. The project positions itself not as a monolithic coding AI, but as a composable framework that organizations customize to their specific workflows. Its core philosophy — that tool curation matters more than quantity — reflects lessons learned from production deployments of coding agents at scale. ## Architecture and Design Open SWE is built on LangGraph and the Deep Agents framework, providing a structured approach to coding agent construction: | Component | Purpose | |-----------|--------| | Invocation Layer | Multi-platform triggers via Slack, Linear, and GitHub (@openswe mentions) | | Sandbox Engine | Isolated cloud environments via Modal, Daytona, Runloop, LangSmith | | Tool System | Curated toolset: shell execution, URL fetching, HTTP requests, PR creation | | Context Engine | AGENTS.md files, issue/thread history, and code context gathering | | Orchestrator | Subagent spawning and deterministic middleware for task coordination | | Validation | Safety nets and automated code quality checks before PR submission | | Messaging | Real-time status updates during task execution | The architecture maps directly to the internal systems built at leading engineering firms. LangChain provides detailed comparisons showing how each architectural decision — invocation, sandboxing, tool selection, context gathering, orchestration, and validation — aligns with or diverges from approaches taken by Stripe, Ramp, and Coinbase. ## Key Capabilities **Multi-Platform Invocation**: Agents can be triggered from Slack messages, Linear tickets, or GitHub issues and PRs by mentioning `@openswe`. This meets developers where they already work, reducing friction for adopting AI-assisted coding workflows. **Isolated Cloud Sandboxes**: Every task runs in its own isolated environment, supporting Modal, Daytona, Runloop, and LangSmith as sandbox providers. This ensures that agent activity cannot accidentally affect production code or interfere with other tasks. **Curated Tool Philosophy**: Rather than exposing hundreds of tools, Open SWE provides a carefully selected set covering shell execution, web fetching, HTTP requests, PR creation, and platform-specific commenting. The framework argues this curated approach yields better agent performance than broader but noisier tool sets. **Context Engineering**: The framework gathers context from AGENTS.md files (similar to CLAUDE.md), issue/thread history, and code context to provide agents with the information they need. This structured approach to context reduces hallucination and improves task relevance. **Parallel Task Execution**: Multiple coding tasks can run simultaneously in separate sandboxes, enabling teams to process backlogs of issues or feature requests concurrently. **Subagent Orchestration**: Complex tasks can be decomposed into subtasks handled by specialized subagents, with deterministic middleware ensuring coordination and consistency. **Real-Time Communication**: Agents provide status updates during execution, so developers can monitor progress and intervene if needed through the same platform (Slack, GitHub) where the task was initiated. ## Getting Started Open SWE provides Docker-ready deployment with comprehensive documentation: ```bash # Clone and configure git clone https://github.com/langchain-ai/open-swe.git cd open-swe cp .env.example .env # Configure sandbox provider and LLM API keys # Deploy with Docker docker compose up ``` Organizations can then connect their Slack workspace, Linear project, or GitHub repositories to start triggering coding tasks through natural language messages. ## Limitations Open SWE requires significant infrastructure setup including sandbox providers, LLM API access, and platform integrations. The framework is Python-centric, which may not align with all engineering stacks. Self-hosting introduces operational complexity compared to managed solutions. While the curated tool approach improves reliability, it can limit agents' ability to handle tasks requiring specialized tooling. The framework is relatively new and its production track record outside of LangChain's own deployments is still developing. Complex tasks involving multiple repositories or cross-service changes may exceed current orchestration capabilities. ## Who Should Use This Open SWE is designed for engineering teams that want to build and own their coding agent infrastructure rather than relying on third-party SaaS products. Organizations with existing LangChain/LangGraph investments can integrate it seamlessly. Teams processing high volumes of routine coding tasks — bug fixes, dependency updates, code migrations — benefit most from the autonomous execution model. Companies with strong security requirements appreciate the self-hosted, sandboxed architecture. Engineering leaders studying how firms like Stripe and Ramp build internal AI tools will find Open SWE's architectural documentation uniquely valuable.