OpenAI report says coding agents sped up science software builds

OpenAI has published a field report examining eight scientific computing projects where coding agents helped reduce development and runtime work. The projects covered areas such as genomics, immunology, statistics, and RNA sequencing, with Codex used in most cases and Claude Code used in some. The report highlights tasks such as packaging cleanup, performance tuning, backend migrations, and language ports. Examples include faster runtimes in tools like HI.SIM and hifiasm, a TensorFlowtoPyTorch migration in MHCflurry, and Rustbased rebuilds that improved speed and maintainability. Across the case studies, the main takeaway was that agents were effective at narrowly defined implementation work, but humans still had to verify scientific correctness, build acceptance tests, and decide whether rebuilt tools should remain compatible with older outputs. The report suggests that verification and stewardship remain the key challenges, even as AI reduces the cost of software engineering work.