Coding agents are helping scientific teams modernise software that often survives long after the paper or grant that created it. An OpenAI field report describes eight genomics, immunology, statistics and sequencing projects where agents improved builds, ports or runtime.
The examples are contributor case studies published by a vendor whose product appears in every project. Their most consistent finding is still important: generating code has become easier, while establishing scientific equivalence remains stubbornly human work.
Agents can tackle neglected engineering debt
Academic tools are commonly maintained by small teams without dedicated engineers. Packaging becomes obsolete, performance bugs persist and a dependency change can make valuable research software difficult to install.
The report used Codex alone in five projects and Codex with Claude Code in three. The cyvcf2 genomics library replaced a legacy build system with a unified process. MHCflurry moved from TensorFlow and Keras to PyTorch while maintaining compatibility with released model weights.
HI.SIM received two largely autonomous optimisation passes that reportedly reduced runtime by 31% across a representative test set without changing output. Hifiasm improved 25% on its optimisation target and roughly 15% on separate human-sequencing data.
Agents were effective when a direct reference implementation or exact output could serve as a judge. They set up benchmarks, proposed changes and handled repetitive migration work. Human contributors still supplied profiling, redirected repeated failures and decided whether an apparently successful change made domain sense.
This pattern lowers the threshold for maintenance that is necessary but difficult to fund. It does not turn a non-specialist into the scientific owner of a rebuilt tool.
Large rewrites show both speed and fragmentation risk
A Rust port of statistical models from the R package bayesm matched estimates within a defined tolerance and ran 2.3 to 2.7 times faster on one thread and 4.4 to 9.5 times faster on eight. Extensions requiring statistical judgement needed direct human validation.
Agents also helped recreate STAR, a 20,000-line RNA-sequence aligner that had lost active maintenance. The work became feasible in weeks, but a person still reviewed more than 900 plots before release because the model could confidently approve incorrect output.
RustQC consolidated 15 quality-control tools into one program and reportedly achieved 60-times faster runtime and 25-times lower disk input/output. Related rebuilds of FastQC and Trim Galore ran seven and three times faster. HelixForge, a GPU-native mutation simulator, also reported about a 60-times speedup.
Fast rebuilding can fragment science. If several laboratories create subtly incompatible replacements, results become harder to compare over time. A performance improvement that alters numerical behaviour or file conventions may damage reproducibility even when tests pass.
Some changes returned to original upstream projects, while the STAR replacement moved to new community stewardship because its predecessor was abandoned. Ownership decisions are as important as implementation choices.
Verification and stewardship are the real constraints
Across the case studies, agents produced strong first drafts and then consumed human time on edge cases, numerical discrepancies and acceptance tests. They could not reliably judge whether their own output was scientifically valid.
Teams need reference datasets, exact parity checks, predefined tolerances and simulated examples with known answers before an agent begins. Performance benchmarks should run alongside correctness, determinism and resource-use tests.
Every port also requires a maintenance plan: who reviews issues, releases security fixes, documents differences and decides when the reference implementation changes? Without that commitment, cheaper code generation may create more abandoned repositories.
The report does not establish a general productivity rate for science or independently verify every claimed speedup. It does show a credible use case where clear specifications and existing behaviour provide strong feedback.
Coding agents can rescue valuable software from technical debt and make ambitious ports affordable to small teams. Their success should be measured not when code compiles, but when scientists can reproduce trusted results and know who will care for the tool next year. In research computing, stewardship is part of correctness.
