OpenAI has released a new field report tracking eight scientific computing projects where coding agents cut runtimes.

The report documents projects that used Codex on its own in five cases and a combination of Codex and Anthropic’s Claude Code in three others. Worth flagging upfront: this is a vendor publishing a survey of its own product’s application in research settings, built from case studies written by the contributors involved.

That doesn’t make the underlying pattern less worth examining. Research software has a documented maintenance problem. Tools built to accompany a single paper, coded by small academic teams without dedicated engineering support, tend to accumulate technical debt that nobody has the budget or mandate to pay down.

OpenAI’s report argues agents can address that debt, and the eight projects it cites span genomics, immunology, statistics, and RNA sequencing.

What tasks the agents undertook

The tasks split roughly into three categories: packaging and build-system cleanup, performance optimisation on existing code, and full language or backend ports.

cyvcf2, a Python library for reading genomic variant files, had its legacy build and packaging system replaced with a newer, unified process, according to contributor Brent Pedersen, who noted that going fast with agents is one thing, but going far in science still needs “expert guidance, understanding, taste, and care.”

HI.SIM, a DNA-sequencing read simulator, saw two largely autonomous optimisation passes from GPT-5.2 and GPT-5.6 that contributor Andrew Ho says cut runtime by 31 percent across a representative test set without altering output.

Ho, who describes himself as neither a genomics specialist nor a C programmer, called the outcome “nothing short of magical” from an end-user perspective, having previously lost time to performance bugs and packaging problems he could recognise but not personally fix.

Hifiasm, used for genome assembly from PacBio HiFi reads, got a 25 percent runtime cut on its optimisation target and roughly 15 percent on separate human sequencing data, per contributor Suyash Shringarpure.

Shringarpure described the agent setting up its own benchmark scaffolding and proposing candidates independently, though he stressed that supplying profiling results and steering the model away from repeated failure modes remained work only a human could do.

MHCflurry, which predicts protein fragments presented to T cells, had its TensorFlow/Keras backend migrated to PyTorch while keeping compatibility with previously released model weights, a change contributors Alex Rubinsteyn, Sergey Feldman, and Timothy O’Donnell frame as the kind of “unglamorous, labour-intensive upkeep” that keeps open-source scientific projects alive rather than left to decay.

bayesm-rs, a Rust port of statistical models from R’s bayesm package, matched the original software’s estimates within a pre-set tolerance and ran 2.3–2.7 times faster on a single processor thread, climbing to 4.4–9.5 times faster across eight threads. According to contributors Andrew Bai and Andrew Ho, the agents handled anything with a direct reference to check against quickly and correctly; extensions requiring statistical judgement the original code never pinned down needed direct human validation instead.

Rust ports and a GPU redesign push the pattern further

Three further projects – rustar-aligner, svb, and kuva – involved Rust builds carried out with coding agents, including a full recreation of STAR, a widely used RNA-sequence alignment tool that had lost active maintenance.

Contributor James M. Ferguson says agents change what’s worth attempting: rewriting a 20,000-line aligner by hand isn’t a sensible use of time, but with an agent it becomes weeks of steered work. Verification, he added, is a separate matter entirely. A model can claim a plot looks fine, but checking over 900 of them by eye before release still fell to a person.

RustQC consolidated 15 separate RNA-sequencing quality-control tools into a single program that contributor Phil Ewels says cut runtime by 60 times and disk input/output by 25 times, with companion rebuilds FastQC-Rust and Trim Galore running seven and three times faster respectively while preserving the original tools’ behaviour.

Ewels also flagged the downside: cheap rebuilds bring their own risk, because tools that diverge in behaviour fragment the community and make results from different labs incomparable over time. “The technology is the easy part,” he said. “Stewardship is the open question.”

HelixForge, a GPU-native rebuild of the mutation-simulation tool BAMSurgeon, reportedly cut runtime by around 60 times on a benchmark involving real human data, according to contributors Mamad Ahangari, Varun Goyal, and Hassan Masoudi, who also say it produced mutation frequencies closer to requested targets and resolved several bugs that generated artefacts in the original tool.

Verification, not code generation, is the constraint now

What comes through across all write-ups is that agents handled well-scoped implementation requests capably but couldn’t judge whether their own output was scientifically sound.

Contributors describe agents expressing confidence in work that contained clear errors, which pushed the actual burden onto humans to build acceptance tests: exact output matching, parity checks against an existing tool, or answers established beforehand using simulated data.

Projects tended to proceed in stages, with agents producing fast first drafts and the remaining time going into edge cases and small numerical discrepancies that a benchmark alone wouldn’t catch.

Lower engineering costs cut both ways. They let a two-person team take on a rebuild that would once have needed a grant-funded engineering hire, and they make it easier for three different labs to produce three incompatible versions of the same tool. Changes to MHCflurry and cyvcf2 went back into their original upstream projects. rustar-aligner moved to new community stewardship because the tool it replaced had already been abandoned.

The OpenAI report points toward a specific choice rather than a general endorsement: decide who owns a rebuilt tool, and secure that commitment, before the first line of agent-generated code ships.