How Agentic AI Is Transforming Scientific Software Development

Problem

Scientific software that underlies genomics, protein research, physics simulation and data processing is often maintained by a small number of long‑standing developers. As languages, hardware and computational demands evolve, legacy codebases become bottlenecks and rewriting them requires substantial engineering effort.

Agentic AI approach

OpenAI’s report "Scientific Computing in the Age of Agentic AI" examined whether autonomous coding agents can alleviate this burden. Eight case studies were documented—five using Codex alone and three combining Codex with Claude Code—covering routine maintenance, targeted optimizations, large‑scale language migration and GPU‑native redesign.

Case study: rustar‑aligner

The RNA‑seq aligner STAR, a >20 000‑line C++ program, was reimplemented in Rust. Because scientific correctness requires exact reproducibility, the team did not rely on automatic translation; instead agents generated code from scratch while researchers defined the architecture, testing strategy and validation criteria. The resulting rustar‑aligner achieved tie‑adjusted parity of 99.815 % on single‑end RNA‑seq tests and 99.883 % on paired‑end tests, and its suffix‑array output matched the original byte for byte.

Case study: SVB nanopore codec

A nanopore data compression codec was reimplemented, delivering 1.7–2.9× speed‑ups over the most widely used existing tools. The VBZ2 double‑strand decoder was about 2.32× faster than the traditional three‑stage pipeline.

Recurring themes

Agents perform well on concrete, narrowly scoped requests but cannot reliably assess scientific validity; they may produce confident yet incorrect results.

Human reviewers must validate outputs using external references, measurable acceptance standards, or simulated data with known answers.

Successful projects adopt an iterative, feedback‑driven workflow: split the overall goal into small tasks, benchmark intermediate results, and apply human judgment to resolve edge cases and numerical discrepancies.

Long‑term management

Scientific software encodes undocumented conventions, compatibility constraints and user trust that simple code translation cannot capture. In the reported projects, changes to MHCflurry and cyvcf2 were merged back into their upstream repositories, while rustar‑aligner, after the original STAR project was abandoned, was handed over to a new community for stewardship. Coordination with existing maintainers is advised; when independent rewrites are necessary, clear ownership and a credible maintenance plan are required to avoid creating another abandoned code base.

Overall insight

Coding agents can rewrite hard‑to‑maintain tools, explore optimizations unattainable by traditional methods, extend functionality, and shorten the gap from idea to usable software. Nonetheless, decisive human input on what to build, how to validate results, and who will maintain the code remains the critical factor for sustainable scientific infrastructure.

Report link: https://cdn.openai.com/pdf/scientific-computing-in-the-age-of-agentic-ai-an-exploratory-field-report.pdf

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
科学家如何利用编码智能体来现代化科研软件。