Open SWE is an open-source framework designed to empower organizations to build their own internal coding agents. Inspired by the practices of elite engineering teams at companies like Stripe, Ramp, and Coinbase, Open SWE provides a readily customizable architecture for automating software development tasks. The primary goal of Open SWE is to streamline engineering workflows, reduce human oversight, and integrate seamlessly with existing communication and project management tools.

At its core, Open SWE leverages the power of Large Language Models (LLMs) to automate tasks such as code generation, bug fixing, and feature implementation. It is built upon the LangGraph and Deep Agents frameworks, providing a robust and flexible foundation for building sophisticated coding agents. The framework offers a comprehensive set of features, including the ability to trigger tasks from Slack, Linear, and GitHub, providing a convenient and integrated user experience.

One of the key features of Open SWE is its use of isolated cloud sandboxes. Each task runs within its own remote Linux environment, ensuring that any potential errors or unintended consequences are contained. This approach prioritizes safety and prevents the agent from making changes directly to production systems. Open SWE supports multiple sandbox providers, including Modal, Daytona, Runloop, and LangSmith, allowing users to choose the best option for their needs.

Open SWE emphasizes tool curation over tool quantity. It provides a focused set of tools, including the ability to execute shell commands, fetch web pages, make API calls, commit and open pull requests, and interact with Linear and Slack. This curated approach ensures that the agent has the necessary capabilities while minimizing complexity and potential security risks. The framework also incorporates context engineering, gathering information from both an `AGENTS.md` file (allowing for repo-specific instructions) and the full context of the triggering issue or thread.

The orchestration within Open SWE is handled through a combination of subagents and middleware. Subagents enable the main agent to spawn child agents for parallel tasks, while middleware provides deterministic hooks that run around the agent loop. This allows for features such as injecting follow-up messages, ensuring PR creation, and gracefully handling tool errors. The framework also supports multiple invocation methods, including Slack, Linear, and GitHub, allowing engineers to interact with the agent through their preferred channels.

Open SWE is designed to be highly customizable. Users can swap out the sandbox, model, tools, triggers, system prompt, and middleware to tailor the agent to their specific needs and workflows. The framework also includes built-in GitHub OAuth for authentication and automatically opens draft pull requests upon completion of a task. The agent is designed to run linters, formatters, and tests before committing changes, and a PR safety net ensures that a PR is opened even if the agent doesn't do it itself.

In essence, Open SWE provides a powerful and flexible framework for building internal coding agents. It offers a combination of features, including isolated sandboxes, curated tools, context engineering, and robust orchestration, all designed to streamline engineering workflows and improve productivity. The framework's open-source nature and emphasis on customization make it an attractive option for organizations looking to automate software development tasks and integrate AI-powered agents into their engineering processes.