AI coding agents hit 60x speedups on legacy science code, but can't tell if the results are garbage

A new field report from OpenAI and several academic partners confirms what many research software developers have suspected: AI coding agents are startlingly good at modernizing creaky, decades-old scientific codebases — and alarmingly bad at knowing whether they’ve broken the science in the process.

The paper documents eight case studies, mostly in biology, where tools like Claude Code and Codex were put to work. In the most dramatic result, the RustQC project combined 15 separate quality-control tools into a single program. Runtime on a large dataset plummeted from 15 hours and 34 minutes to just 14 minutes and 54 seconds. That’s a speedup of more than 60 times. Another project, HelixForge, rebuilt a genomic data generator for GPUs and saw a 98.6x speedup in its main compute step. The pattern is clear: for well-defined, grunt-level programming tasks — porting 10,000 lines from TensorFlow to PyTorch, replacing a build system, rewriting C++ in Rust — these agents are now genuinely useful.

The catch is verification. The agents present their work with total confidence, even when it’s wrong. Philip Ewels, who led RustQC, describes them as “eloquent, convincing, and confidently wrong in ways that are easy to miss.” He never let the models judge their own output, instead building an independent test harness. The bayesm case study was particularly insidious. A Rust rewrite ran 2-20x faster than the original, but two advanced methods contained bugs that were nearly invisible from the output alone. One involved an inverted control parameter that silently used the reciprocal of the intended values. Researchers only caught it by running thousands of synthetic datasets with known results.

The division of labor that emerged across all eight projects was consistent: humans defined the goals, success criteria, and validation tests; agents wrote the code. This isn’t the death of the research programmer. It’s a role shift. The bottleneck moves from writing code to designing the tests that prove the code is correct — a skill set that’s harder to teach and, for now, uniquely human. An earlier attempt to port MHCflurry to PyTorch failed in early 2025 with older models. Developer Sergey Feldman pins that failure on model capability, not the approach, suggesting the threshold for reliability is very recent and still rising.