From Execution to Discovery : Physical Verification as the Key Gap in AI Agents for Science

Why can AI agents operate reliably in digital environments such as software development yet still struggle to achieve autonomous scientific discovery? We argue that this gap reflects not only limitations in model capability but also differences in how agent capabilities are externalized and how reliably environments can verify the consequences of agent actions. In this survey, we trace the technical evolution of AI agents and propose a three-stage framework of capability externalization: reasoning externalization makes knowledge, memory, and intermediate inferences inspectable; execution externalization makes tool use and long-horizon workflows governable; and world-interaction externalization integrates environmental evidence into hypothesis revision and action selection. Materials discovery provides a demanding testbed for all three stages, spanning literature-grounded hypothesis generation, computational design, automated experimentation, and iterative knowledge updating. Although agents can increasingly complete predefined research workflows, workflow automation alone does not constitute scientific discovery. Sustained autonomy requires agents to assess whether new observations support or challenge a hypothesis and to select subsequent experiments accordingly. We therefore identify environmental verifiability as a central systems constraint on agent autonomy: the extent to which an environment can translate actions into timely, attributable, reproducible, and affordable evidence. This perspective positions physical verification capacity as a key scaling dimension for autonomous scientific discovery.