Directory / SWE-agent
SWE-agent
Verified
Princeton’s agent that made SWE-bench a product category.
SWE-agent is an open-source coding agent from Princeton NLP, designed around the SWE-bench task: given a GitHub issue, produce a patch. It exposes an agent-computer interface (ACI) for editing and navigating repos. It is research software that people still run, not a hosted IDE.
- Model independence
- Multi-model
- Deployment
- Self-hosted
- Open source
- Yes
- License
- MIT
- Pricing model
- Free
- Added / updated
- 19 Aug 2026 / 19 Aug 2026
swe-agent.com · Docs · Source
Editorial take
SWE-agent is why half the category quotes SWE-bench. Use it as a harness and a baseline, not as a team’s daily driver. Mini-SWE-agent and later forks exist because the original ACI was a research interface. If your factory metric is “SWE-bench %,” you are measuring the bench.
Strengths
- Public MIT source and a documented agent-computer interface.
- Native fit for SWE-bench-style issue-to-patch evals.
- Model-agnostic; you bring the LM.
Limitations
- Research UX. Not a product with SSO and audit logs.
- Optimized around a benchmark distribution that is not your monorepo.
- No hosted control plane.