SWE-agent

SWE-agent is a Princeton/Stanford research project that takes a GitHub issue and automatically attempts to fix it using your LLM of choice. Achieves over 74% on SWE-bench verified. The recommended successor, Mini-SWE-Agent, is just ~100 lines of code and requires no special tool-calling interface — it operates purely through bash commands. Works with any model via litellm and supports Docker/podman for sandboxed execution.