Deep Agents Open Source Open SWE: An Open-Source Framework for Internal Coding Agents
Explore Open SWE to prototype internal coding agents with isolated sandboxes and curated toolsets.
Integrate Open SWE into your internal agent pipeline by setting up a sandbox backend and adding the curated toolset.
Summary
Open SWE is a new open‑source framework released by the LangChain team on 17 March 2026 that captures the architectural patterns observed in internal coding agents built by Stripe, Ramp, and Coinbase. The framework is built on Deep Agents and LangGraph and provides a composable agent harness that can be upgraded as Deep Agents improves, while keeping org‑specific tools, prompts, and workflows in configuration. Open SWE introduces an isolated sandbox backend that supports Modal, Daytona, Runloop, and LangSmith, with each conversation thread receiving a persistent sandbox that is automatically recreated if it becomes unreachable. The curated toolset includes shell commands, fetch_url, http_request, commit_and_open_pr, linear_comment, slack_thread_reply, and the native Deep Agents tools such as read_file, write_file, edit_file, ls, glob, grep, write_todos, and task for subagent spawning. Context engineering is achieved through an AGENTS.md file that injects repository‑wide conventions into the system prompt, and by pulling full Linear issue or Slack thread history as source context before the agent starts. Orchestration combines deterministic middleware hooks like check_message_queue_before_model and open_pr_if_needed with subagent spawning via the task tool, ensuring reliable execution and automatic PR creation when needed. Invocation is handled through Slack, Linear, and GitHub, with the agent replying in‑thread and posting PR links, while validation is enforced by running linters, formatters, and tests before committing, with a backstop middleware that opens a PR if the agent finishes without one.
Key changes
- Open SWE framework built on Deep Agents and LangGraph
- Agent harness composition enables upgrade path without rebuilding customizations
- Sandbox backend supports Modal, Daytona, Runloop, and LangSmith with persistent threads
- Curated toolset includes shell, fetch_url, http_request, commit_and_open_pr, linear_comment, slack_thread_reply, and Deep Agents core tools
- Context engineering via AGENTS.md and source context from Linear or Slack
- Deterministic middleware hooks (check_message_queue_before_model, open_pr_if_needed) combined with subagent spawning
- Invocation through Slack, Linear, and GitHub with in‑thread status updates
- Validation layer runs linters, formatters, and tests before committing, with automatic PR creation if needed