AI coding agents are rewriting genomics software, but the last 10% still takes human grit

The messy underbelly of scientific software is no secret to anyone who’s actually worked in a lab. Tools built by overworked grad students to get a paper out the door often ship with brittle build systems and zero documentation. A new field report from OpenAI, covering eight agent-assisted projects in the life sciences, suggests that’s starting to change — and the implications go way beyond just cleaner code.

Contributors across the board reported a dramatic shift in their daily work. Instead of grinding through implementation details, researchers described their role evolving into what one called “verification and orchestration.” They’d specify what to build and define how to measure correctness, while agents handled the tedious plumbing. In one striking example, a team used GPT‑5.5 to completely modernize the build and packaging system for cyvcf2, a widely used library for parsing genomic variant files. This kind of maintenance work typically languishes because no one has the time or funding to do it. The agent just… did it.

But the report is refreshingly honest about the limits. Agents were fast and confident, but also frequently wrong with a straight face. They couldn’t reliably judge whether their output was scientifically valid. The teams that succeeded didn’t just trust the AI — they built rigorous external validation into their workflow. One project used exact output agreement against an existing tool as a target. Another leaned on simulated data with answers established in advance. The winning pattern was clear: break the goal into small, feedback-driven chunks, and never skip the step where a human with deep domain expertise signs off.

As Brent Pedersen, one of the contributors, put it: “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.” The bottleneck isn’t engineering labor anymore. It’s human judgment. The “last mile” of resolving subtle numerical edge cases still ate up the most time, proving that the hard part of science isn’t generating code — it’s knowing when the code is actually right.