From Execution to Discovery: Physical Verification as the Key Gap in AI Agents for Science
Why do AI agents succeed in digital environments yet struggle with autonomous scientific discovery? This review proposes three stages of capability externalization: reasoning, execution, and world interaction. Using materials discovery as a testbed, it identifies environmental verifiability as a central constraint on sustained autonomy: actions must yield timely, attributable, reproducible, and affordable evidence. Physical verification infrastructure is therefore essential to connecting automated workflows with hypothesis revision and reproducible discovery.


































