Beyond Static Agents: A Six-Dimensional Taxonomy and Survey of Self-Evolving LLM Agents for Healthcare

Abstract

Large language model (LLM)-based agents are increasingly being explored for healthcare tasks such as clinical decision support, care coordination and autonomous workflow execution. Beyond static pipelines, recent systems claim to self-evolve by adapting their tools, memory, reasoning, policy, context and coordination strategies over time. However, what constitutes true self-evolution in healthcare agents remains unclear and is inconsistently evaluated. In this work, we introduce a six-dimensional taxonomy to rigorously characterize self-evolving LLM agents in healthcare, comprising Tools & Capabilities, Memory & Knowledge, Context & Prompt, Policy & Strategy, Multi-Agent Coordination and Reasoning. Using this taxonomy, we systematically evaluate 18 representative healthcare agent papers under strict mechanism-level criteria that separate genuine self-evolution from surface-level performance improvements. Our results reveal that memory evolution is the most common capability with over half of the surveyed systems exhibiting reflective or distilled memory growth, while only 5.6% demonstrates memory refactoring. In contrast, prompt self-revision appears in 33.3% of the papers and tool-selection evolution appears in only 5.6%. Multi-agent architectures (i.e., systems with multiple interacting agent roles/modules) are frequent, yet learned coordination remains limited, with only 16.7% exhibiting evolving protocols or topologies. We find no evidence of formal policy optimization or reasoning-mechanism evolution across these 18 papers. Overall, our analysis shows a clear gap between claimed adaptivity and verifiable self-evolution in current healthcare agents. This taxonomy provides a principled foundation for evaluating future self-evolving LLM systems and supports more rigorous, reproducible progress toward adaptive clinical AI.