Opportunities for AutoML in the Agentic Era

Automated Machine Learning (AutoML) reduces manual effort in model and pipeline design but still suffers from high search cost and limited ability to encode domain knowledge. This survey examines the convergence of AutoML with large language models (LLMs) and AI agents and frames it as bidirectional enhancement: LLMs can steer AutoML search through reasoning and prior knowledge, while AutoML optimizes LLMs across pre-training, post-training, and test-time. Agents enable natural-language, end-to-end ML pipeline automation, and AutoML can tune multi-agent roles, prompts, and communication protocols. We further highlight opportunities in AI4Science, including automated scientific model design, workflow automation, and interpretable pattern extraction. We synthesize empirical findings and outline open challenges in reliability, cost, scalability, and validation, offering practical guidelines for practitioners.