mini-swe-agent
A roughly 100-line AI software engineering agent that solves GitHub issues or assists on the command line, scoring above 74% on SWE-bench verified with no tool-calling interface.
mini-swe-agent is a minimal AI software engineering agent built around the idea of asking how simple an agent can be while still working nearly as well as more complex ones. The agent class itself is about 100 lines of Python, with a bit more code for the environment, model and run script, and no dependency beyond bash: it does not use a tool-calling interface at all, so it can run with literally any model.
It scores above 74% on the SWE-bench verified benchmark, is used by organizations including Meta, NVIDIA, Essential AI, IBM, Nebius and Anyscale, and starts much faster than larger coding agents. Its execution history is completely linear -- every step just appends to the messages, so there is no difference between the trajectory and what is sent to the model, which the project notes is useful for debugging and fine-tuning.
It deploys to local environments, Docker, Podman, Singularity/Apptainer, bubblewrap and other sandboxes, executing each action as an independent subprocess.run call, and supports any model through litellm, OpenRouter or Portkey.