Technology RadarTechnology Radar
Assess

Full deep dive: Open SWE Architecture Breakdown

Open SWE is LangChain's open-source framework for building autonomous coding agents — a multi-agent system (Manager/Planner/Programmer/Reviewer) built on LangGraph and Deep Agents, with pluggable sandbox isolation, Slack/Linear/GitHub invocation, and human-in-the-loop plan review. MIT-licensed, 8.7K+ GitHub stars, directly inspired by the patterns Stripe Minions, Ramp Inspect, and Coinbase converged on independently.

Architecture Overview

Open SWE implements seven core architectural components:

  1. Agent Harness — built on Deep Agents and LangGraph, providing the multi-agent orchestration
  2. Sandbox Environment — isolated cloud execution per task (Modal, Daytona, Runloop, LangSmith)
  3. Tool Suite — ~15 curated tools covering shell, web, Git, and integrations
  4. Context Engineering — AGENTS.md files plus source context injection
  5. Orchestration — subagent spawning and middleware hooks
  6. Invocation — Slack, Linear, and GitHub integration points
  7. Validation — prompt-driven checks with safety-net PR creation

Why It's in Assess

The architectural patterns are proven (Stripe, Ramp, Coinbase all converged on them), and this is the first open-source implementation. However, Open SWE launched March 17, 2026 and has no independent production adoption data yet. Study the patterns — especially the multi-agent pipeline and sandbox isolation model — but assess maturity before deploying internally.

Key Characteristics

Property Value
Provider LangChain Inc.
Architecture Multi-agent (Manager → Planner → Programmer + Reviewer)
Built on LangGraph, Deep Agents
Sandbox providers Modal, Daytona, Runloop, LangSmith (pluggable)
Toolset ~15 curated tools
Default model Claude Opus 4.6 (model-agnostic)
License MIT (GitHub)
Inspired by Stripe Minions, Ramp Inspect, Coinbase internal agents
Sources Blog, GitHub