Open SWE: LangChain's Open-Source Framework Distills the Internal Coding Agent Architecture That Stripe, Ramp & Coinbase Built Independently
There's a compelling pattern hiding in plain sight across the engineering organizations at Stripe, Ramp, and Coinbase: each company built its own internal AI coding agent — Minions, Inspect, and Cloudbot respectively — entirely independently, and yet all three arrived at nearly the same architecture. Isolated cloud sandboxes. A curated toolset of roughly 15 tools. Subagent orchestration for parallel workstreams. Deep integration with Slack, Linear, and GitHub. LangChain noticed the convergence and decided to package what these teams discovered the hard way into an open-source framework called Open SWE.
Built on LangGraph and Deep Agents, Open SWE gives any engineering organization a production-tested starting point rather than a blank canvas. The underlying premise is sound: if three highly capable, independently operating teams converged on the same design choices under real production pressure, those choices are probably the right defaults. The framework captures the architecture — not just the tooling — which is what tends to be hardest to get right when building agents that need to operate reliably over time across messy, real-world codebases.
This matters beyond the fintech world. The convergence story validates a set of architectural primitives that have been debated heavily in the agentic coding space: how many tools is too many, when to fan out with subagents, how to keep context manageable. Open SWE turns those lessons into code any team can fork, customize, and deploy — dramatically lowering the barrier to standing up a serious internal coding agent without the multi-quarter investment these companies made.