Mini SWE-Agent is a minimal but practical AI coding agent developed by the Princeton and Stanford teams behind SWE-bench and SWE-agent. Its core claim is that: after SWE-bench and SWE-agent drove the coding agent revolution in 2024, agents in 2025 can be 100 times simpler while maintaining comparable performance.
Purpose and Positioning Mini SWE-Agent is intended to be a hackable tool, not a black box. It is designed to be simple enough to understand at a glance, convenient enough for daily workflows, and flexible and extensible. It is especially suitable as a fast command-line tool for local use, for agent development requiring minimal control flow, and for scenarios involving fine-tuning (FT) or reinforcement learning (RL) while avoiding overfitting to a specific agent scaffold. Officially, it is recommended to make mini-swe-agent the default choice.
Main Capabilities The agent is implemented in only about 100 lines of Python (about 100 lines in total including env, model, and script), with no fancy dependency libraries. It provides no tools other than bash, and does not even use the language model's tool-calling interface, instead focusing entirely on enabling the language model to make full use of the shell. This means it can run with any model and requires no software packages to be installed when running in a sandbox environment, only bash. It scores over 74% on the SWE-bench verified benchmark and starts far faster than Claude Code. Officially, it is mentioned that Gemini 3 Pro reached a 74% score on SWE-bench verified through mini-swe-agent on November 19, 2025. In addition, mini-swe-agent has also been used for Ramp SWE-Bench and defeated Claude Code and Codex on DeepSWE.
Usage and Deployment Mini SWE-Agent supports multiple deployment methods such as local environments, docker/podman, singularity/apptainer, bubblewrap, and contree. It supports all models through litellm, openrouter, portkey, etc., and supports the /completion and /response endpoint, interleaved thinking, and other features. Users can get started by reading the official documentation's "Get started with mini-SWE-agent" and "Learn to use the mini command," and can set up and customize their own agent. The project's current version is v2, and the official documentation provides a migration guide from v1; v1's documentation and branch are still available.
Architecture and design features Mini SWE-Agent executes actions via subprocess.run, and each action is completely independent rather than maintaining a stateful shell session. This makes executing actions in a sandbox very simple (just replace subprocess.run with docker exec), and also easy to extend. Its history is completely linear, with each step simply appending content to the messages, so there is no distinction between the trajectory and the messages passed to the language model, which is very helpful for debugging and fine-tuning.
Use cases If you want a fast command-line tool that can run locally, an agent with very simple control flow, faster, simpler, and more stable sandboxing and benchmarking, or if you are doing fine-tuning or reinforcement learning and do not want to overfit to a specific agent scaffold, mini-swe-agent is a suitable choice. If you want to experiment with different tool sets (each tool having its own interface) or different history processors, you should use swe-agent. Both provide excellent SWE-Bench performance and a trajectory browser.
Limitations and known information The official materials do not provide a clear pricing model, price, platform restrictions, or specific installation commands. The features mentioned in the documentation are based on the official descriptions, and information that could not be determined has not been provided for now.
