Breakthroughs in large language models are accelerating the evolution of medical artificial intelligence from standalone models to multi-agent systems. To organize this landscape, we propose a hierarchical taxonomy spanning Level 0 to Level 5. Given that foundational Level 0 to Level 2 architectures face cognitive and collaborative bottlenecks in complex clinical workflows, this survey focuses primarily on the architectural paradigm shift between Level 3 and Level 4. Specifically, we examine how systems transition from static Level 3 collaboration based on predefined roles to Level 4 self-evolving architectures that iteratively refine internal strategies through interactive feedback. Concurrently, we explore governance frameworks required to ensure system safety and trustworthiness under stringent clinical standards. Finally, we highlight that the capacity to dynamically generate agents and reconstruct collaborative topologies based on real-world scenarios will be essential for surpassing current capability limits. Building on this concept, we envision a Level 5 medical AI landscape defined by evolutionary agent swarms. To help researchers continuously track this field, our updated GitHub repository (https://github.com/Tiany1Yu/Medical_Multi_Agent_Systems_Survey_Papers) is publicly available.

Large language models (LLMs) have shown great potential in multi-disciplinary team (MDT) medical consultations. However, long, multi-round, multi-role interaction trajectories inevitably lead to severe information dilution and context window overload, triggering context collapse which destabilizes reasoning. Furthermore, prior systems typically rely on unstructured trajectory history storage without structurally distilling key information or reflecting on errors, severely limiting continuous learning capabilities. We propose MDTeamGPT, a context-resilient and self-evolving multi-agent framework. Mechanistically, we introduce a specialized Lead Physician mechanism combined with a Residual Context architecture to compress and reorganize multi-round consensus, effectively mitigating context overload and reducing computational costs. For memory, we design a Dual Knowledge Base system comprising a CorrectKB for verified trajectories and a ChainKB for reflective error analysis, enabling self-evolution via retrieval from both successes and failures. We evaluated our framework on standard text datasets (MedQA, PubMedQA), multimodal benchmarks (VQA-RAD, SLAKE), and collected more complex clinical problems. Experimental results show that MDTeamGPT substantially outperforms existing baselines across both text-based and multimodal tasks, while also demonstrating superior diagnostic performance and stability in complex clinical scenarios.