A 15-Hour Pipeline Now Runs in 15 Minutes. 7 of 8 Teams Still Had to Prove It.
OpenAI's new field report has eight case studies of coding agents rewriting real scientific software. The speedups are enormous. The agents still couldn't tell anyone when they were done.
NeuroX AI · August 3, 2026
OpenAI and academic partners published a field report on agentic coding in scientific software on August 1, with eight case studies from real labs. The headline number holds up: a genomics QC pipeline that took 15 hours 34 minutes now finishes in 14 minutes 54 seconds — over 60x, on a 186-million-read human dataset, with disk traffic cut from 2.5 TB to 0.1 TB.
It wasn't a one-off. HelixForge, a GPU-native rewrite, ran 59.6x faster end-to-end and 98.6x on its main compute stage. Another team ported MHCflurry from TensorFlow to PyTorch across roughly 10,000 lines and 130 files while keeping previously trained weights loadable.
Then the part worth reading twice. In seven of the eight cases the contributors remained the "principal adjudicators of success," and the report is blunt that agent self-assessments "did not provide reliable evidence of completion." Humans picked the reference datasets, fixed the tolerances in advance, and investigated every discrepancy. Developing the validation framework was itself a substantial part of the work.
That matches what we see on every production build. More agent autonomy didn't remove human judgment — it moved it upstream into specification and validation design. The last mile still ate the most time: edge cases, numerical drift, failures that only show up on realistic workloads.
If the agent's own "done" is your only gate, you don't have a system. You have a prototype running 60x faster.