mini swe agent

Report Abuse

Basic Information

mini-swe-agent is a compact open-source AI agent implemented in roughly 100 lines of Python that automates developer workflows, notably resolving GitHub issues and executing shell-based tasks. It is intended for researchers who need a minimal, auditable baseline for benchmarking, fine-tuning, or reinforcement learning and for developers who want a small, readable command-line tool for everyday automation. The project emphasizes a model-centric design with no custom tools beyond bash and a fully linear message history that makes debugging and prompt inspection straightforward. It includes a simple CLI called mini, optional visual mode, Python bindings, batch inference utilities, and a trajectory browser. The repository documents deployment options including local execution and container runtimes and provides installation methods via pip, pipx, or from source.

Links

Categorization

App Details

Features
The repo is intentionally minimal: the core agent logic is about 100 lines of Python and relies only on bash for external actions. Every agent action is executed via subprocess.run and the agent maintains a completely linear message history, which simplifies reproducibility and debugging. It is model-agnostic and thus can be run with a variety of language models. The project reports strong benchmark performance on SWE-bench (68% GitHub issue resolution in the referenced benchmark). Supporting features include a simple CLI (mini), a visual UI mode, batch inference, a trajectory browser, Python bindings, container-friendly deployment (docker, podman, singularity, apptainer), and test coverage indicators.
Use Cases
mini-swe-agent serves as a lightweight, hackable baseline for research and practical automation. For researchers it provides a stable, small scaffold for benchmarking LMs, fine-tuning and RL experiments without introducing complex tool interfaces. For developers it offers a fast local CLI to automate repo tasks, reproduce agent behavior, and inspect agent trajectories. For engineers it is trivial to sandbox and deploy because actions are independent subprocess calls, making it easy to swap execution backends or run inside containers. Its simplicity reduces risk of overfitting to an agent scaffold, speeds up debugging and prompt engineering, and integrates into batch or CI workflows via Python bindings and CLI.

Please fill the required fields*