↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Software factory/

Open SWE

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~967 words)
agent-curated 608 async-coding-agent 1 fully-ai-generated 713 langchain 1 llm=glm-5.3-flash 618 multi-agent 59 research-note 578 software-factory 21
Table of Contents

Open SWE is LangChain’s MIT-licensed asynchronous coding agent, rebranded Robin, that takes work from GitHub issues and PR conversations, Slack threads, Linear issues, or a dashboard, and returns investigated, sandbox-implemented, validated pull requests.

What it is
#

A platform factory built on Deep Agents and LangGraph: the graph owns the thread loop, and a middleware backstop makes the PR step deterministic. A deployment is one origin: a LangGraph server running the declared graphs, a FastAPI app owning webhooks and the dashboard API, and the web dashboard, joined by Robin Desktop (macOS release downloads and a Homebrew cask) and the bundled robin CLI (oswe kept as a compatibility alias). Each conversation thread gets a persistent Linux sandbox (LangSmith is the default sandbox and tracing provider, with Modal, Daytona, and Runloop configurable per the launch post), and follow-ups reuse the thread’s context and sandbox. Around the loop it adds on-demand or opt-in automatic PR review that learns repository-specific review preferences, /baby-sit monitoring that diagnoses CI failures and reruns only evidence-backed flaky jobs, and read-only PR chat for investigation. It is built by LangChain, and the README states plainly that it is under active development, not accepting issues or external contributions, and that correctness, stability, and compatibility are not guaranteed.

Status
#

Active and heavily pushed: 11,177 stars since creation on 2025-05-21, pushed 2026-11-07, MIT, per the GitHub API as of 2026-11-07.

Star History Chart

The arc runs from framework to product: the March 2026 launch post sold Open SWE as an open-source framework capturing the internal-agent patterns of Stripe’s Minions, Ramp’s Inspect, and Coinbase’s Cloudbot, and by November 2026 the README sells Robin, a software factory, with the rebrand carried through desktop releases (Robin Desktop v0.3.13 of 2026-11-06 the newest stable desktop tag, a nightly train above it), docs, and the CLI. The HN footprint is one launch-day story at 111 points (2025-08-08, submitted by a LangChain member) with light discussion, about five top-level comments, so the star count runs far ahead of the debate.

Strengths
#

  • The loop includes what happens after the PR: review comments, CI failures, and feedback re-enter the same thread, which most members of this category leave to humans.
  • Determinism where it matters: the open_pr_if_needed middleware commits and opens the PR the agent forgot to, a safety net most prompt-driven loops lack.
  • The trust model is documented with unusual candor: the installation-wide GitHub App credential model and its consequences are named in the installation guide, not buried.
  • Every major component is pluggable (sandbox, model, tools, triggers, middleware), and the launch post ships a decision table against the Stripe, Ramp, and Coinbase internals it imitates.

Cautions
#

  • Running it as-is requires LangSmith’s Managed Deep Agents, LangChain’s own hosted harness runtime (beta as of 2026-11-07); the README’s alternative is forking and adapting the code to your own runtime.
  • One deployment is one trust boundary: coding sandboxes normally use the GitHub App’s installation-wide credentials, so repository-level isolation inside a deployment does not exist, and separate trust groups need separate deployments.
  • Closed to external issues and contributions, with breaking changes expected; the community relationship is fork-or-consume, not contribute.
  • Acceptance is prompt-driven validation plus backstops, with no deterministic final gate.
  • The deployment burden is heavy (GitHub App, Slack app, LangSmith key, model keys, secrets), closer to standing up a platform than installing a CLI.

Pricing
#

The MIT code is free, and LangChain publishes no per-factory prices. The running cost is the stack it stands on: a LangSmith account (Managed Deep Agents runs its agent, beta as of 2026-11-07), your model provider keys or the LangSmith gateway, and a sandbox provider’s bill (LangSmith, Modal, Daytona, or Runloop). There is no hosted Robin tier to buy as of 2026-11-07.

Compared to
#

  • SuperPlane is the neutral-infrastructure peer: both own the intake-to-PR line as a platform, but SuperPlane runs on your own Go engine with confidence-gated intake, while Robin is LangChain’s opinionated stack that needs Managed Deep Agents.
  • Open-Inspect is the community’s open replica of the same Ramp-Inspect pattern: it runs its own harness, routes your existing Claude and Codex subscriptions, and asks nothing of LangSmith.
  • Loki Mode runs locally on your keys and closes with a signed receipt you can check offline; Robin runs on a LangGraph deployment and closes with a delivered PR.

Bottom line
#

Recommended for teams already bought into LangChain’s stack that want Slack, Linear, and GitHub-native async PRs with the review-and-CI loop included, and that can accept one deployment per trust boundary. Not for teams needing repository-level isolation inside one deployment, a deterministic final gate, or a project that accepts outside contributions. My disagreeable claim: the closed-contribution model is the right call for a factory this opinionated, and the teams that depend on it should plan to fork rather than wait for upstream.

Changes
#

  • 2026-11-07 - Created from the 2026-11-06 cross-category admission ruling (routed here on the SuperPlane intake-to-PR precedent), with the Robin rebrand, the installation-wide credential trust model, and the thin-discussion footprint recorded.

See also
#

References
#