Agentic AI for Scientific Discovery: Progress, Challenges, and Future Directions (arxiv.org)

This survey synthesizes recent work on “agentic” AI systems designed to accelerate scientific discovery—models that combine reasoning, planning, and autonomous decision-making to perform literature review, generate hypotheses, design and run experiments, and analyze results. The authors categorize existing tools and architectures deployed across chemistry, biology, and materials science, and compile the implementation frameworks, common datasets, and evaluation metrics currently used to judge agentic systems’ performance. By mapping the landscape, the paper clarifies where agentic approaches have already enabled closed-loop workflows and where gaps remain. For the AI/ML community, the survey highlights both opportunity and responsibility: agentic systems can scale scientific throughput and enable reproducible, automated experimentation, but they also expose technical challenges in reliability, calibration, and benchmarked evaluation. Key implications include the need for standardized datasets and metrics tailored to discovery tasks, tighter integration between simulators and real-world labs, robust uncertainty estimation and failure modes, and design patterns that support human-AI collaboration. The authors call for research focused on trustworthy calibration, interpretability, and ethical governance to ensure agentic tools augment rather than mislead scientific practice—setting a roadmap for building safer, more effective discovery agents.