SWE-Agent

Princeton and Stanford researchers built this agent around a purpose-built agent-computer interface for language models: a windowed file viewer, a directory-wide search command, and an edit command guarded by a linter that rejects syntactically broken changes, all running in a Docker sandbox through the companion SWE-ReX runtime. Point sweagent run at a GitHub issue and repository, pick a model (hosted APIs such as Anthropic or OpenAI, or local models), set a per-instance spending cap, and it attempts a patch; run-batch evaluates whole suites such as SWE-bench. A single YAML config governs behavior, which suits agent research and ablations, and command-line and web inspectors replay saved trajectories. EnIGMA, a capture-the-flag security mode, still needs the older 0.7 branch. Installation is a git clone plus an editable pip install, with Docker recommended and cloud execution as a fallback. Code is MIT licensed. The team says its much smaller successor, mini-swe-agent, has superseded it and now gets most development effort, so new projects should start there and this listing best serves researchers reproducing published results.