SWE-agent/mini-swe-agent

What it solves

mini-swe-agent is a minimal AI software engineering agent designed to be a hackable, high-performance tool rather than a complex black box. It removes the overhead of specialized tools and complex interfaces, allowing the language model to interact directly with the system shell to solve software engineering tasks.

How it works

Unlike traditional agents, it uses a radically simplified architecture:

• Bash-only interaction: It does not use tool-calling interfaces; it relies entirely on the language model's ability to use bash commands to perform actions.
• Linear history: Every step is simply appended to the message history, making the trajectory easy to debug and fine-tune.
• Independent execution: Actions are executed via subprocess.run, meaning each action is independent rather than maintaining a stateful shell session, which improves stability and simplifies sandboxing.
• Broad compatibility: It supports various models via LiteLLM, OpenRouter, and Portkey, and can be deployed in local environments or sandboxes like Docker, Podman, and Singularity.

Who it’s for

• Developers who want a fast, simple command-line tool for AI-assisted coding.
• Researchers doing fine-tuning (FT) or reinforcement learning (RL) who need a baseline system that doesn't overfit to a specific agent scaffold.
• Users seeking a stable and fast way to run benchmark evaluations on SWE-bench.

Highlights

• Minimalist design: The core agent class is roughly 100 lines of Python.
• High performance: Scores over 74% on the SWE-bench verified benchmark.
• Flexible deployment: Supports local environments, Docker, Podman, Singularity, and more.
• Model agnostic: Works with any model that supports completion or response endpoints.
• Trajectory browser: Includes a tool to visualize the agent's steps.