OpenAI Report Links Coding Agents to Faster Science Software Builds

OpenAI has released a new field report tracking eight scientific computing projects where coding agents successfully reduced execution runtimes. The report documents projects that relied on Codex on its own in five instances and a combination of Codex and Anthropic's Claude Code in three others. While worth noting upfront that this is a vendor publishing survey results of its own product applied in research settings based on case studies from contributors, the underlying patterns remain highly revealing.

Research software notoriously suffers from a documented maintenance problem. Tools built for a single academic paper by small teams without dedicated engineering support tend to accumulate technical debt that lacks a dedicated budget to clear. OpenAI's report argues that coding agents can effectively tackle this debt across genomics, immunology, statistics, and RNA sequencing fields.

The assigned tasks generally split into three distinct categories:

• Packaging and build-system cleanup
• Performance optimisation on existing code
• Full language or backend ports

Stock photo for illustration only, not from the actual event

For example, cyvcf2, a Python library for reading genomic variant files, had its legacy build and packaging system replaced with a newer, unified process. Contributor Brent Pedersen noted that moving fast with agents is one thing, but advancing further in science still requires expert guidance, understanding, taste, and care.

Similarly, HI.SIM, a DNA-sequencing read simulator, underwent two largely autonomous optimisation passes executed by GPT-5.2 and GPT-5.6. Contributor Andrew Ho stated that this cut runtimes by 31 percent across a test set without altering the output. Ho, describing himself as neither a genomics specialist nor a C programmer, called the result nothing short of magical from an end-user perspective after previously struggling with performance and packaging bugs.

Hifiasm, utilized for genome assembly from PacBio HiFi reads, achieved a 25 percent runtime reduction on its core optimization target and roughly 15 percent on separate human sequencing data, according to contributor Suyash Shringarpure. Shringarpure explained that the agent set up its own benchmark scaffolding independently, though profiling and steering the model away from repeated failures remained strictly human work.

"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."

James M. Ferguson

Three additional projects—rustar-aligner, svb, and kuva—involved Rust builds executed with coding agents, including a complete recreation of STAR, an unmaintained RNA-sequence alignment tool. Contributor James M. Ferguson remarked that agents fundamentally change what is worth attempting, turning massive rewrites into manageable, steered efforts, though visual verification of over 900 plots still fell entirely to humans.

Stock photo for illustration only, not from the actual event

Furthermore, RustQC consolidated 15 separate RNA-sequencing quality-control tools into a single program, cutting runtime by 60 times and disk I/O by 25 times according to Phil Ewels. Meanwhile, HelixForge, a GPU-native rebuild of BAMSurgeon, also slashed runtime by roughly 60 times on real human benchmarking data while resolving artifact-generating bugs, per contributors Mamad Ahangari, Varun Goyal, and Hassan Masoudi.

Source: AI News