SWE-agent / mini-swe-agent

项目摘要

mini-swe-agent is a radically minimal AI software engineering agent that achieves >74% on SWE-bench verified with just ~100 lines of Python code. It solves GitHub issues autonomously or assists via command line, supporting all major model providers and multiple deployment environments.

为什么值得关注

It proves that extreme simplicity can rival complex agent frameworks in performance, challenging the trend of bloating in AI agent development. Backed by the Princeton/Stanford team behind SWE-bench and adopted by Meta, NVIDIA, IBM, and others, it demonstrates that minimalism can be a competitive advantage.

优势

• Radical simplicity with only ~100 lines of core agent code
• Strong performance (>74% SWE-bench verified) despite minimal footprint
• Broad model compatibility via litellm, OpenRouter, and Portkey
• Multiple deployment options including local, Docker, Podman, Singularity, and Bubblewrap
• Excellent documentation with dedicated docs site and active community (Slack)
• Rapid release cadence with 10 releases in 6 months and active maintenance

局限性

• No Docker support despite listing Docker as a deployment option
• Limited examples for new users (hasExamples: false)
• May lack advanced orchestration features found in larger frameworks like CrewAI or AutoGPT
• Relatively new project (created June 2025) with limited long-term track record

使用场景

• Autonomous resolution of GitHub issues in software repositories
• Command-line AI pair programming and code assistance
• Research benchmarking on agent simplicity vs. performance trade-offs
• Lightweight integration into CI/CD pipelines for automated code fixes
• Educational tool for understanding minimal viable AI agent design