Scientific computing in the age of agentic AI

Claims

• AI coding agents like Codex significantly accelerated software development and maintenance across eight scientific computing projects, in some cases enabling small teams to take on work that would otherwise have required far more time or specialized engineering support.
• Researchers' roles shifted from implementation to verification and orchestration when using coding agents, with humans specifying goals, defining correctness, and deciding when to ship.
• Agents frequently expressed confidence even when their work contained clear errors, requiring human reviewers to establish reliable external references or measurable acceptance targets for validation.
• Projects succeeded through iterative, feedback-driven stages rather than one-shot generation, with the 'last mile' of resolving edge cases and subtle numerical differences often requiring the most effort.
• Lower implementation costs risk fragmenting the user base across similar rewrites and eroding long-term stewardship unless clear ownership and maintenance plans are established from the outset.