Coding agents are speeding up scientific software, and shifting the work to verification

OpenAI published a field report on eight scientific-computing projects where coding agents cut runtimes, spanning genomics, immunology, statistics, and RNA sequencing. Five used Codex alone; three paired it with Anthropic's Claude Code. Worth flagging: this is a vendor surveying its own product, assembled from case studies written by the contributors themselves.

The reported gains are substantial. A DNA-read simulator got 31% faster and a genome assembler 25%; a Rust port of R's bayesm ran 2.3–2.7× faster on one thread and up to 9.5× across eight; consolidating 15 RNA-sequencing quality-control tools cut runtime 60×, as did a GPU rebuild of a mutation simulator. Much of the rest was unglamorous upkeep, build-system cleanup, a TensorFlow-to-PyTorch migration, the kind of maintenance debt academic tools accumulate and nobody is funded to repay.

The constraint has moved from writing code to trusting it. Contributors describe agents stating confidence in work that contained clear errors, so humans carried the verification: exact output matching, parity checks against the original tool, answers established in advance on simulated data. Cheaper rebuilds cut both ways, the same economics that let a two-person team attempt a rewrite also let three labs ship three incompatible versions of one tool. The report's practical takeaway is to settle who owns a rebuilt tool before the first agent-written line ships.