AI coding agents can help modernize neglected research software, but they do not remove the hardest part of the job: proving the science still works. A field report from OpenAI and academic partners describes eight case studies, mostly in biology, using tools such as Codex and Claude Code.
The projects ranged from build-system updates to major rewrites. In one case, agents helped port MHCflurry, an immunology model, from TensorFlow to PyTorch. Another project rebuilt the STAR genome aligner in Rust, then compared results on 10,000 yeast sequencing reads.
Some gains were large. RustQC combined 15 quality-control tools and cut runtime from more than 15 hours to under 15 minutes on a large dataset. HelixForge, a GPU-based replacement for a synthetic genomics tool, ran parts of its pipeline nearly 100 times faster.
The limitation is confidence. The report says agents completed well-defined tasks quickly but could present wrong work convincingly. For scientific software, the bottleneck shifts from writing code to designing rigorous checks.