Mini SWE-Agent

Mini SWE-Agent is an AI coding agent, built by the team behind SWE-bench and SWE-agent with roots at Princeton and Stanford, that solves software engineering tasks autonomously by executing bash commands to understand codebases, make changes, and run tests. Its core agent is just 100 lines of Python with no complex dependencies, yet achieves over 74% accuracy on the SWE-bench verified benchmark. It runs across local systems, Docker, Singularity, and Bubblewrap sandboxed environments, works with any LLM through litellm, OpenRouter, and Portkey integrations, and uses only bash with no specialized tool-calling interfaces, appending each step linearly to message history for straightforward debugging. It is adopted by Meta, NVIDIA, and IBM.