Scientific computing in the age of agentic AI

Scientific computing relies on software tools that have struggled to keep pace with modern data generation, often resulting in fragile infrastructure that requires constant maintenance and limits research progress. Coding agents are beginning to modernize this landscape by handling tedious engineering tasks, allowing researchers to spend less time on implementation and more time on scientific discovery and validation. A field report examined eight scientific computing projects where agents assisted with software maintenance, optimization, and redesign, finding that agents effectively handled specific tasks but still required human experts to verify results and judge scientific validity. The report identifies a shift where researchers move from implementation work to defining goals, validating outputs, and maintaining long-term stewardship of tools, with the lasting value depending on human decisions about what to build and who will support it over time.