mini-SWE-agent

A minimal coding agent from the SWE-agent team that gives the model nothing but bash (SWE-agent).

The agent class is about 100 lines of Python. It doesn't use the model's tool-calling interface at all. The model writes a bash command, and each one runs independently with subprocess.run, so swapping in docker exec is enough to sandbox it. The history is strictly linear, so the messages the model sees are the trajectory itself. Its README claims over 74% on SWE-bench Verified.

It's one of the eleven harnesses compared in Harness Engineering Anatomy, Architecture, and Evolution of Coding Agents, and the shared backbone for every method in the Context Language Models experiments. See also Thin Harnesses.