SWE-agent

SWE-agent taught a generation of agents how to talk to a repository, and reading it is still the fastest way to understand why every tool on this board looks the way it does. You hand it a GitHub issue, it works the fix in isolation, and any model you hold a key for takes the attempt. The daily trait that decides it is that there is no daily: it is a harness for experiments, with no editor and no memory of yesterday.

Pick it as a citable research scaffold and a teaching tool. Pick OpenHands for daily work, and mini-swe-agent for the same authors' current thinking in a hundred lines.

The risk is that nobody is home. The maintainers state the project is in maintenance-only mode and point to mini-swe-agent as the successor, so an issue filed here is an issue filed against history, and a fix you need is a fix you write. That is an honest status and a terminal one.

Run it for a paper reproduction, not for work, and pin the commit you reproduced against. What it does right: every run is sandboxed in Docker, so a wrong command breaks a container rather than a checkout, which is more than most commercial agents on this board offer.

SWE-agent is the reference implementation of the agent-computer interface: the interface, not the model, is the object of study, and the paper measured that. The documented figure, 12.29 percent resolved on the full SWE-bench test set, is from 2024 and scoped correctly to the full set rather than a subset. Later 1.0 results are described as "state of the art" with no published number on the docs.

A claim without a figure is not a result, and a figure without a subset is not comparable. The observation: the project that defined the benchmark discipline now declines to follow it.

No funding round, no pricing page, no revenue: SWE-agent is a research project from Princeton and Stanford under an open license, and the 18-month question is about the fork tree, not the cap table. Adoption signal is real: 20,200 GitHub stars and a benchmark half the category cites in its marketing, which is distribution the universities never intended to monetize.

Nobody acquires a research scaffold; they hire the authors, and the authors have already moved to their next project. Position: cite it, do not deploy it, and watch where the people go.

The demo is a patch out of a container. There is no vendor to sign a data-processing agreement with; the maintainers are two universities, and a university does not answer a security questionnaire. No SSO, no audit log, no support line, no retention policy beyond whatever the model provider offers. Cost is tokens only, with per-instance and total limits configurable, which is the one governance feature present.

CI fit exists in theory, since it runs headless, and nobody on staff would own it in practice. Not yet, and given the project's status, not later either.

The cleanest thing on the board to read: MIT, every prompt in the source, and a config format small enough to hold in your head. LiteLLM means my local OpenAI-compatible server is one config value away, so the whole loop runs on my hardware with nothing leaving the room. No MCP, which I forgive in a research harness that predates the protocol.

The code stopping under me is a feature: I would rather own a frozen MIT harness than rent a moving proprietary one, and a fork of something small and finished is a fork I can actually maintain.