mini-SWE-agent is a minimalist, high-performance AI agent designed for automated software engineering tasks. Created by the team behind SWE-bench and the original SWE-agent, it represents a shift toward simplicity, focusing on effectiveness through a streamlined design. The agent consists of roughly 100 lines of Python, intentionally omitting complex abstractions and specialized interfaces in favor of a direct, bash-based interaction model. It is widely adopted by organizations including Meta, NVIDIA, and various leading universities for research and practical development tasks.

Functionality-wise, the agent performs as a command-line utility that interacts with local or sandboxed environments using only the shell. By bypassing the traditional tool-calling interfaces of language models, it maintains a completely linear conversation history, which simplifies debugging, fine-tuning, and performance analysis. Every action is executed via an independent subshell, ensuring high stability and making it easy to scale across various compute environments, including Docker, Singularity, and Bubblewrap.

Some of the key features are:

• Minimalist Architecture: Highly compact codebase that requires no complex dependencies.
• Universal Compatibility: Supports virtually any language model through integrations like LiteLLM, OpenRouter, and Portkey.
• Bash-Centric Workflow: Relies entirely on the shell for operations, allowing the model to utilize full Linux capabilities instead of limited, pre-defined tools.
• High Performance: Achieves competitive results on benchmarks such as SWE-bench Verified, outperforming many more complex systems.
• Flexible Deployment: Compatible with multiple sandboxing environments including local, Docker, and Singularity backends.
• Linear Trajectory: Simple, append-only message history makes the agent trivial to audit, debug, and use for fine-tuning.
• Interactive Inspection: Includes an integrated trajectory browser for reviewing agent decision-making processes.

The agent is used by invoking the mini command-line tool, which provides a REPL-like environment. It supports various modes, including a confirmation mode for human oversight and a YOLO (you-only-live-once) mode for automated execution. The agent is highly extensible, allowing users to define custom configurations in YAML files or utilize built-in Python bindings to build advanced agent logic or custom benchmarks.

Some common use cases include:

• Automated Bug Fixing: Resolving GitHub issues by analyzing repositories and applying code patches.
• Software Engineering Research: Acting as a baseline system for developing and evaluating new AI agent architectures.
• Automated Reverse Engineering: Solving tasks within the ProgramBench benchmark by generating original code from binary behavior observation.
• Rapid Prototyping: Generating and testing code for new features or algorithmic implementations within local projects.
• Fine-tuning Data Generation: Creating high-quality agent interaction trajectories to train smaller or specialized models.