MiniMax M2.7: The self-refactoring Agent architecture

MiniMax's M2.7 introduces a self-refactoring agent architecture where the model autonomously rewrites its own scaffold — the tools, skills, memory, and workflow rules it operates within — without any retraining. After each task, M2.7 analyzes failures, plans and applies changes to its harness, evaluates results, and stores self-criticism in memory for future iterations. Over 100 internal rounds, it discovered optimizations like tuning sampling parameters, writing workflow-specific guidelines, and adding loop detection, achieving a 30% performance improvement. On MLE Bench Lite, it averaged a 66.6% medal rate across 22 ML competitions, tying with Gemini 3.1. The key insight is that model weights never change — only the surrounding system improves, enabling continuous production-time optimization without gradient updates.