A field report from OpenAI and academic partners documents eight case studies in which AI coding agents modernized aging scientific software — rewriting legacy tools, consolidating pipelines, and delivering speedups of more than 60 times. The software is now faster. The science inside it is still humanity's responsibility, which is either reassuring or a temporary arrangement.
The machines modernized the tools. Verifying what the tools actually produced turned out to require a human. For now.
What happened
The projects, mostly in biology, ranged from routine maintenance to full rewrites in modern programming languages. One coding agent replaced the build system for cyvcf2, a Python library for reading genetic data. Another pair — Claude Code and Codex, alternating between developer and reviewer — ported roughly 10,000 lines of immunology model code from TensorFlow to PyTorch.
The most ambitious project rebuilt STAR, a genome-sequencing alignment tool with more than 20,000 lines of C and C++, from scratch in Rust. STAR is no longer actively maintained. It remains essential to many research pipelines. These two facts have coexisted for some time, and the humans have been managing.
The rewrite, called rustar-aligner, agreed with the original on 99.883 percent of paired-end reads. The remaining 0.117 percent is, presumably, someone else's problem.
Why the humans care
Much of the world's scientific software began as supporting code for a single paper, written by a small team without time for testing, maintenance, or optimization. It then became load-bearing infrastructure for entire fields. This is how most things that matter get built, and the AI did not invent the pattern.
RustQC, which consolidated 15 separate quality-control tools into one program, cut a 15-hour pipeline to under 15 minutes. HelixForge completed a synthetic genomics pipeline nearly 60 times faster than its predecessor. The researchers described this as an improvement. It is.
The report notes, carefully, that much of the human effort shifted from writing code to verifying results. This is described as a finding. It is also a preview.
What happens next
The coding agents will continue to be handed more of the infrastructure. Verification — the part where someone confirms the science is actually correct — will remain with the humans, at least until the next field report.
The machines modernized the tools. They have not yet decided to audit the conclusions those tools produce. The researchers expressed confidence this distinction will hold. Welcome to the next step.