Self Evolving Live SWE Agent (github.com)

The Live-SWE-agent has been introduced as the first live, runtime self-evolving software engineering agent that can autonomously expand and adapt its capabilities while tackling real-world software engineering tasks. This innovation is significant for the AI/ML community as it challenges the traditional reliance on proprietary, engineered scaffolds for LLM-based benchmarks. With scores reaching 79.2% on SWE-bench Verified for Claude Opus 4.5 and 77.4% for Gemini 3 Pro, the Live-SWE-agent demonstrates a clear ability to outperform existing models in a fair and transparent manner. By providing a minimal and open framework that allows for apples-to-apples comparisons amongst different AI models, Live-SWE-agent streamlines the evaluation process and potentially accelerates advancements in software engineering automation. Built on the popular mini-swe-agent framework, it emphasizes flexibility and adaptability, showcasing a state-of-the-art solve rate of 45.8% on SWE-Bench Pro. The release of Live-SWE-agent 1.0.0 invites further community engagement, encouraging the submission of model evaluation results to enhance benchmarking and foster an environment of robust, equitable development in AI software engineering.