Decades of scientific software sit unmaintained because the graduate student who wrote it graduated.

Research code rots quietly until it breaks at the worst moment

Small academic teams ship tools alongside papers with no dedicated engineering support, and the technical debt piles up with no mandate to clear it. The result is genomics pipelines running on legacy build systems, performance bugs that researchers can identify but cannot personally fix, and tools that quietly fall out of use.

Eight projects, three task types, one consistent pattern

OpenAI report links coding agents to faster science software builds is a field report from OpenAI tracking eight scientific computing projects where coding agents handled packaging cleanup, performance optimisation, and full language ports across genomics, immunology, statistics, and RNA sequencing. Researchers pointed the agents at specific codebases, reviewed the output, and in five cases used Codex alone; in three others they combined Codex with Anthropic’s Claude Code. The most striking single result came from HI.SIM, a DNA-sequencing simulator, where two largely autonomous optimisation passes cut runtime by 31 percent without altering output.

The researchers who felt this most were not the ones writing the code

• Bioinformaticians maintaining legacy Python libraries who need build system modernisation without regression risk.
• Computational scientists who can spot a performance bottleneck in C but lack the language depth to fix it themselves.
• Research software engineers responsible for tools accompanying published papers that still get citations years after the original team dissolved.

HI.SIM contributor Andrew Ho described himself as neither a genomics specialist nor a C programmer, and called the 31 percent runtime reduction he got from the agent passes nothing short of magical from where he was standing.

The vendor-report caveat is real, but the underlying problem is older than OpenAI

This is OpenAI publishing case studies written by contributors using its own products, which means the eight projects are not a neutral sample. What makes the pattern worth tracking is that research software maintenance has been a documented, underfunded problem for years, and coding agents are now cheap enough to run on it at scale, which puts pressure on every scientific computing team still treating refactoring as a luxury.

What teams are running agents on right now

• Replace legacy build and packaging systems in active Python libraries.
• Port performance-critical code from Python to Rust or C without specialist staff.
• Run autonomous optimisation passes on simulation code and benchmark the delta.
• Clear dependency debt on tools tied to published papers still in active use.

Pricing for Codex access runs through the OpenAI API with usage-based billing; Claude Code is billed separately through Anthropic.

The report’s own contributors flag the hard limit: going fast with agents is possible, but going far in science still requires expert guidance, taste, and domain knowledge the agent does not have.

GitHub Copilot Workspace targets similar refactoring workflows but is aimed at product engineering teams rather than research contexts. Google’s Jules agent is moving into the same agentic coding space with a focus on asynchronous background tasks.

Coding agents are moving into scientific infrastructure faster than most labs have a policy for it

The shift from autocomplete to autonomous multi-pass optimisation on real research codebases is happening whether institutions are ready or not. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.