What Open SWE Is
TL;DR LangChain releases Open SWE, an open-source framework for building autonomous coding agents in the enterprise. Inspired by Stripe, Ramp, and Coinbase internal systems, it provides a composable architecture with isolated sandboxes, ~15 curated tools, and Slack/GitHub/Linear integration. MIT license, 6,200+ stars.
On March 17, 2026, LangChain released Open SWE, an open-source framework for deploying autonomous coding agents. An Open SWE agent can analyze a codebase, plan an implementation, write code, run tests, review its own work, and open a Pull Request - all asynchronously.
The project came from a concrete observation: Stripe, Ramp, and Coinbase each independently built very similar systems.
- Stripe developed Minions: ~500 curated tools per agent, AWS EC2 sandboxes
- Ramp built Inspect: composed on OpenCode, Modal containers, visual DOM verification
- Coinbase created Cloudbot: three operational modes, auto-merge capabilities
All three share the same architectural choices: isolated execution, curated tools, Slack/GitHub/Linear integration, sub-agent orchestration. Open SWE extracts these common patterns into a reusable framework.
The 7-Component Architecture
Open SWE is built on LangGraph and the Deep Agents framework.
1. Agent Harness
Open SWE composes on top of Deep Agents rather than forking it. When the underlying framework evolves, customizations are preserved. Same approach as Ramp with OpenCode.
2. Isolated Sandbox
Each task executes in a dedicated cloud Linux environment with full shell access. Supported providers: Modal, Daytona, Runloop, LangSmith. Custom backend possible. If the agent breaks something, the blast radius is contained to its sandbox.
3. Curated Tools
Unlike Stripe's ~500 tools, Open SWE provides about fifteen:
execute- shell commandsfetch_url- web scrapinghttp_request- API callscommit_and_open_pr- Git operationslinear_comment,slack_thread_reply- communication- Deep Agents built-ins: file operations, search, sub-agent spawning
4. Context Engineering
An AGENTS.md file at the repo level (same principle as CLAUDE.md for Claude Code) injects conventions and architectural context into the system prompt. Combined with the full issue or thread history, the agent understands the project without a discovery phase.
5. Orchestration and Sub-agents
The main agent can spawn sub-agents in parallel. Deterministic middleware guarantees certain behaviors:
check_message_queue_before_model: live message injection during executionopen_pr_if_needed: guarantees PR creation even if the LLM forgetsToolErrorMiddleware: error handling
6. Multi-Surface Invocation
The agent triggers via existing tools:
- Slack: mention with
repo:owner/namesyntax - Linear: comment on a ticket
- GitHub: feedback on a PR
7. Validation
Prompt-driven validation (the agent tests its own code) + deterministic middleware that guarantees critical steps execute regardless of LLM behavior.
What This Changes for Developers
With an agent like Open SWE, the developer workflow shifts. Instead of picking up a ticket, reading the code, implementing, testing, and opening a PR, the developer assigns the ticket to the agent via Slack, validates the proposed execution plan, then reviews the generated PR. Mechanical implementation is delegated. Specification, supervision, and review become the core of the work.
This shift creates new skill requirements. Configuring an AGENTS.md that documents project conventions for the agent. Writing tickets precise enough to be processed autonomously. Spotting typical LLM mistakes during code review - API hallucinations, tests that pass without actually testing, unnecessary abstractions.
Developers who master these practices - and understand the underlying patterns like ReAct loops, plan-and-execute, and multi-agent orchestration - delegate effectively. Others spend more time fixing the agent than they would coding themselves.
The Potential for Companies: Code Produced 24/7
An Open SWE agent runs in its cloud sandbox with no schedule. A team can assign it refactoring tickets, minor bug fixes, dependency migrations, or unit test additions during nights and weekends. Monday morning, the PRs are ready for review.
Coinbase pushes the concept further with Cloudbot: their agent auto-merges its PRs when tests pass and automated review is positive. Human developers only intervene on complex cases.
An agent doesn't make architecture decisions or debug subtle performance issues. But it absorbs the repetitive work that takes up a significant portion of development time. It costs a few dollars per task (API + sandbox), runs continuously, and scales horizontally - 10 tickets in parallel, each in its isolated sandbox.
Training Your Teams: The Competitive Edge
A company whose developers can operate a coding agent increases its production capacity. Maintenance tickets, refactoring, tests, and dependency migrations are handled continuously. Developers focus on architecture, product, and complex problems. Time-to-market accelerates.
Conversely, a company that doesn't invest in these skills misses a lever its competitors are already using. A developer who masters agentic coding can deliver at least 10x the output on routine tasks. The velocity gap widens with every sprint.
And deploying an agent without training the team doesn't close that gap. A poorly written AGENTS.md, vague tickets, unreviewed generated code - the agent produces volume but not quality. Training developers to orchestrate and operate AI agents is the prerequisite to extracting real value.
Going Further
Our AI-Augmented Developer training covers in 2 days what teams need to work effectively with agents: Context Engineering, Prompt Engineering, Augmented Code Review. For building autonomous systems based on AI Agents, the AI Engineer training goes deeper over 3 days.
