SWE-agent

Why it's ranked here

This is the project that established the agent-computer interface pattern much of today's coding-agent tooling still uses, from a Princeton and Stanford research team with a peer-reviewed paper behind it. That track record is the case for it, not current momentum: the maintainers themselves now point newcomers toward a smaller successor project and describe this one as superseded. Worth knowing for the ideas and the benchmark work; anyone choosing what to run today should read the maintainers' own recommendation before picking this over the newer one.

What's good

The configuration system is genuinely deep: agent prompts, allowed tools, model choice, and the execution interface are all controlled from YAML files that can be layered and merged rather than edited in code. Command-line usage covers single runs, batch runs, and trajectory replay for debugging, and there are two separate inspectors, a terminal one and a browser-based one, for stepping back through what the agent actually did. Model access goes through a provider-agnostic layer rather than being tied to one vendor's API.

Tradeoffs

The headline caveat is in the project's own documentation: it has been superseded by a simpler successor the maintainers now recommend by default, so anyone adopting this version is opting into the older, more complex codebase over the one actively being improved. The offensive-security mode is currently unsupported on the main line and requires checking out an old release tag to use. This is an academic research codebase first, not a polished product, so expect surface area built for experimentation rather than turnkey deployment.

How to use it well

Reach for this if you want to study or extend the agent-computer interface approach itself, reproduce benchmark results, or need the offensive-security capture-the-flag mode and are willing to pin an older release for it. If you just want a working coding agent to point at a repository today, read the maintainers' own recommendation first: they steer new users toward the newer, simpler successor project instead. Treat the YAML configuration layer as the main lever for adapting it to a new task.

Observed

License

MIT licensed, per both README.md and the pyproject.toml license field.

Packaging

Requires Python 3.11 or newer; packaged as a pip-installable Python project with a sweagent CLI entry point.

CLI commands

CLI subcommands include run, run-batch, run-replay, inspect (terminal), inspector (web-based), quick-stats, merge-preds, traj-to-demo, and remove-unfinished.

Configuration

Agent behavior, prompts, tool access, and model choice are all set through external YAML config files rather than code, and multiple config files can be merged.

Dependencies

Depends on litellm for model access, making the LLM backend swappable rather than fixed to one provider.

Dependencies

Depends on a separate swe-rex package for executing commands in the target environment, with a minimum required version enforced at startup.

Interfaces

Ships a Flask/Flask-SocketIO-based web inspector alongside a Textual-based terminal inspector.

Operating modes

Has two operating modes: a general coding-task/bug-fix mode and a separate offensive-security (capture-the-flag) mode; the latter is documented as unsupported on the current main line and requiring an older release.

Project status

The README states development focus has shifted to a separate, simpler successor project that the maintainers describe as having superseded this one and recommend using instead.