Scientific Computing in the Age of Agentic AI

OpenAI's field report documents eight agent-assisted scientific computing projects, primarily in life sciences, in which AI coding agents (Codex and Claude Code) materially accelerated software development, maintenance, and large-scale language migrations for research teams. Across all cases, agents handled well-scoped implementation tasks effectively but could not self-validate scientific correctness, requiring human researchers to shift from writing code to specifying goals, defining acceptance criteria, and verifying outputs. Projects consistently required iterative, staged workflows rather than one-shot generation, with edge-case resolution and 'last mile' work consuming disproportionate effort. The report flags a structural risk: lower implementation costs may proliferate redundant rewrites and fragment maintainer attention, making long-term stewardship, clear ownership, and maintenance planning as important as the initial productivity gain.