mini-swe-agent - A 100-line AI agent for resolving GitHub issues and command-line support

The Princeton·Stanford team has released mini-swe-agent.

• It's an ultra-lightweight AI software engineering agent built with about 100 lines of Python code.
• It uses only bash as a tool, and can be connected to any model without needing the LLM's tool-calling interface.
• Each action is run independently via subprocess.run, eliminating the need for a stateful shell session.
• It uses a fully linear history structure, so the trajectory and the messages passed to the LLM are identical.
• It achieved over 74% on SWE-bench verified, and reached the same level with Gemini 3 Pro as well.
• It supports various sandbox environments including local, Docker/Podman, Singularity/Apptainer, Bubblewrap, and Contree.
• It can use various models via litellm, openrouter, and portkey, and is compatible with /completion, /response endpoints as well as interleaved thinking.
• It provides the mini CLI, batch inference, a trajectory browser, and Python bindings (DefaultAgent, LitellmModel, LocalEnvironment).
• It's distributed under the MIT license.
• It has reportedly been adopted by Meta, NVIDIA, IBM, Essential AI, Nebius, Anyscale, Princeton·Stanford, and others.