Materials Data & Informatics Laboratory

Since its birth, computational materials science has worked with a clear division of labor. Humans develop algorithms, write codes, and interpret results, while computers crunch numbers. What this division never covered is the work that requires scientific judgment while a calculation is under way, such as recognizing the layered character of a candidate structure and redirecting the search toward a more promising region of configuration space. One line of our work builds agentic workflows that run simulations with language model agents. Paimon is a general platform of this kind, and its central concern is the silent error, a result that looks plausible but is physically wrong, which it suppresses well enough to make such workflows substantially more reliable. The direction we are heading is different in kind. An agentic program is scientific software written for a single clearly bounded responsibility rather than for open-ended research, and that boundedness is what makes rigorous task-specific verification possible, because scope, tools, inputs, outputs, and physical constraints can all be fixed in advance. Modern agent harnesses make such software practical by serving as both a development environment and a runtime for model-based judgment. Delegation is not granted all at once but earned through maturation. Early versions still need frequent human inspection, and each failure a researcher diagnoses is generalized into a permanent construction rule or verification gate, so the range of judgment the program can be trusted with widens with experience. The end state is complete delegation in production, where the program either returns a validated result or reports that no reliable result can be produced, and reaching it is what separates an agentic program from an assi