A field report from OpenAI and academic collaborators has found that AI coding agents can dramatically accelerate the modernization of aging scientific software, rewriting tools in faster languages and slashing runtimes by orders of magnitude. The catch: the same agents cannot reliably judge whether their work is scientifically correct, often presenting buggy code with complete confidence.

The report, published August 1, 2026 and documented by The Decoder, examined eight case studies — mostly in biology — in which research groups used coding agents such as Codex and Claude Code to rescue, optimize, and rewrite critical software. For ongoing reporting on the tools reshaping science and industry, follow our latest AI news.

A Quiet Crisis in Scientific Software

Much of the software underpinning modern research began as supporting code for a single paper. Small academic teams built these tools without the resources for proper testing, maintenance, or optimization. The result is a fragile foundation: programs that remain critical to entire fields but require constant repair. The OpenAI field report suggests AI coding agents could help close that long-standing gap.

The projects ranged from routine maintenance to full rewrites in modern programming languages, and several delivered eye-catching performance gains.

Rebuilding STAR in Rust

One of the more ambitious projects, rustar-aligner, rebuilt STAR from scratch in Rust. STAR is a widely used tool that maps sequencing reads from cells to their corresponding locations in a genome. The original program contains more than 20,000 lines of C and C++ and is no longer actively maintained, even though it remains embedded in many research pipelines.

To verify the rewrite, the team tested both tools on 10,000 short sequencing reads from yeast cells. For single-end reads, rustar-aligner produced the same result as STAR in 99.815 percent of cases. For paired-end reads, the agreement rate was 99.883 percent. The comparison extended beyond mapped locations to include several other key fields both programs produce for each read, and neither tool mapped any reads that the other failed to map.

A 60-Fold Speedup

RustQC delivered the largest speedup by combining 15 separate quality-control tools into a single program. On a large dataset, runtime plummeted from 15 hours and 34 minutes to just 14 minutes and 54 seconds — a speedup of more than 60 times.

Another project, HelixForge, replaced a tool for generating synthetic genomic data with a GPU-accelerated version. In a test using data from one donor and a ten-million-base-pair section of the genome, HelixForge completed the full pipeline 59.6 times faster than the original BamSurgeon, with the main compute step alone running 98.6 times faster.

In a more conventional modernization, GPT-5.5 replaced the outdated build and installation process of cyvcf2, a Python library for reading genetic data. The more involved MHCflurry migration saw Claude Code and Codex alternate between the roles of developer and reviewer while porting roughly 10,000 lines of code from TensorFlow to PyTorch. MHCflurry is an immunology model that predicts which targets immune cells will recognize.

'Confidently Wrong'

The headline finding, however, was a sobering one. Across the case studies, agents completed well-defined tasks quickly but could not reliably judge whether their work was scientifically correct. Even when their code contained errors, the systems often presented it with full confidence.

Brent Pedersen, the developer of cyvcf2, captured the tension in the report: "With coding agents, it's quite easy to go fast; for now, to go far in science, there's still a need for expert guidance, understanding, taste, and care."

Philip Ewels, who led the RustQC project, was blunter. He described the agents as "eloquent, convincing, and confidently wrong in ways that are easy to miss." Ewels never allowed the models to judge the accuracy of their own work, instead building an independent test harness to catch their mistakes.

Errors Hiding in Plain Sight

The bayesm case study illustrated just how hard these errors can be to catch. Its Rust rewrite ran between two and twenty times faster than the original, but the first versions of two advanced methods contained subtle defects. In one, the agent inverted a key control parameter, causing the program to use the reciprocal of the intended values. A separate bug affected the calculation itself. Researchers only discovered both issues after running a detailed calibration test against thousands of synthetic datasets with known results.

The episode underscores a structural shift in how scientists may work alongside AI: rather than writing code themselves, their central task becomes rigorously verifying the results — a role that demands deep domain expertise the agents themselves cannot supply.

What It Means for Science

The findings land at a moment of intense debate over how much autonomy AI systems should be granted in high-stakes domains. The speedups are genuinely transformative — a 60-fold reduction in runtime can turn an overnight job into a coffee break. But the report's most important contribution may be its honesty about the limits. Agents that are fast, fluent, and persuasive, yet unable to self-assess scientific validity, are a powerful tool only in the hands of experts who know exactly what to check.

For the researchers involved, the lesson is clear: AI can rebuild the scaffolding of science at remarkable speed, but human judgment remains load-bearing.

Stay Ahead of AI

From research breakthroughs to enterprise deployments, the pace of change is relentless. Bookmark AI Buzz Wire for ongoing, fact-checked coverage of how artificial intelligence is reshaping science, business, and society.

Read more AI research news →