SWE-agent

Princeton's agent-computer-interface framework for autonomous software engineering (19.8K★), with continuous releases through 2026. The line includes mini-swe-agent, a radically minimal ~100-line agent scoring roughly 65% on SWE-bench Verified and widely used as a baseline harness (5.9K★), and SWE-ReX, the sandboxed, massively parallel code-execution runtime that backs SWE-agent and SWE-smith evaluations. Developed by Press, Yang, Jimenez et al. in Karthik Narasimhan's group.