SWE-agent

SWE-agent

By Princeton NLP

Princeton's agent that autonomously resolves real GitHub issues via an agent–computer interface. Now maintenance-only — superseded by mini-SWE-agent

Maintenance onlyNo updates in 1+ yearv1.1.02025-05-22PythonMIT
19.9k Completely free Open source

Doc version 1.0 · Updated

Related tools

Maintenance-only. SWE-agent has been superseded by mini-SWE-agent — simpler and more flexible while just as performant. New projects should start there; see the FAQ for why.

Overview

SWE-agent gives a language model a purpose-built agent–computer interface (ACI) — commands to view, edit, and test files in a repository — and lets it work through a GitHub issue end to end.
It’s a reference implementation behind many strong SWE-bench results and a clean base for issue-fixing agents.

flowchart LR
  issue[GitHub issue] --> agent[SWE-agent]
  agent -->|ACI commands| repo[(sandboxed repo)]
  repo -->|test results| agent
  agent --> patch[patch to review]

The Code samples tab has the CLI usage.

When to use it

Reach for SWE-agent when the unit of work is a GitHub issue and you want a research-grade, benchmarkable loop focused on automated bug fixing.