100 lines, 74% on SWE-bench
A radically simple 100-line AI agent that solves GitHub issues or aids your command line, scoring over 74% on SWE-bench verified.
The agent class is 100 lines. No tool-calling interface, no framework, just bash. It scores over 74% on SWE-bench verified anyway. Built by the Princeton and Stanford team behind SWE-bench itself. Worth saving as a reminder that most agent complexity is optional.