Open SWE is an open-source framework for building internal coding agents, written in Python and built on LangGraph and Deep Agents. It provides organizations with the architecture used by elite engineering teams at companies like Stripe, Ramp, and Coinbase to create Slackbots, CLIs, and web applications that automate coding tasks with minimal human oversight. The framework is designed to be customizable for individual codebases and workflows while maintaining safety boundaries and proper context management.
The repository is classified across multiple domains including Software Agents, Autonomous Development, Code Generation, Issue Resolution, LLM Agents, Agent Framework, Software Automation, Debugging, Testing, and Code Repair. The repository shares contributors with langchain-ai/langchainjs, langchain-ai/chat-langchain, and openclaw/clawhub.
The core architecture of Open SWE centers on seven key design decisions. First, it composes on the Deep Agents framework rather than forking existing agents, providing an upgrade path while allowing organizational customization. Second, every task runs in isolated cloud sandboxes with full shell access, supporting multiple providers including Modal, Daytona, Runloop, and LangSmith. Sandboxes are persistent per thread and auto-recreate if unreachable, with multiple tasks running in parallel without queuing. Third, the framework uses a curated toolset rather than accumulating tools, including execute for shell commands, fetch_url for web pages, http_request for API calls, linear_comment for ticket updates, and slack_add_reaction and slack_thread_reply for Slack integration. Built-in Deep Agents tools include file operations and subagent spawning via the task tool.
Context engineering in Open SWE relies on AGENTS.md files at repository roots for encoding conventions and architectural decisions, combined with full source context from Linear issues or Slack threads. Orchestration supports subagents for parallel work and deterministic middleware hooks that inject follow-up messages before model calls and notify users when step limits are reached. Invocation occurs through Slack with repo syntax support, Linear via comments, and GitHub through PR comments, with deterministic thread IDs routing follow-up messages to the same running agent. Validation is prompt-driven, with the agent responsible for running linters, formatters, and tests before committing and opening pull requests.
The framework includes optional observability tools for Datadog and LangSmith integration, restricted to authorized users through encrypted credentials stored server-side. Optional Corridor guardrails provide plan analysis capabilities. The comparison table in the documentation shows Open SWE's positioning relative to internal systems at Stripe, Ramp, and Coinbase across dimensions including harness composition, sandbox implementation, tool curation, context gathering, orchestration patterns, invocation surfaces, and validation approaches. The framework supports topics including agents, AI, Anthropic, Claude, LLMs, and OpenAI integrations.
