What was studied
On July 28, 2026, OpenAI published an exploratory field report covering eight agent-assisted scientific-computing projects. Five used Codex alone and three combined Codex with Claude Code. The work ranged from modernizing genomics packaging to language migrations and GPU-native redesigns.
Recurring patterns
- Agents accelerated early implementation, testing, and optimization when tasks were well scoped.
- Researchers shifted from writing every line toward defining goals, designing acceptance tests, reviewing results, and orchestrating iterations.
- The strongest projects used external, measurable criteria such as exact-output agreement, parity with an established tool, expected statistical behavior, or answers fixed in advance with simulated data.
- Edge cases and subtle numerical differences made the last mile slower than the first implementation.
Why it matters
Research software often starts in small academic teams with little time for packaging, testing, optimization, or support. Lower engineering costs could let scientists spend more effort on experiments and interpretation. The same lower cost can also produce fragmented rewrites that divide users and maintenance capacity.
Limits
This is an exploratory set of contributor case studies, not a controlled productivity trial. OpenAI reports that agents could not reliably decide whether outputs were scientifically valid and could sound confident when wrong. Independent tests, expert review, early coordination with upstream maintainers, and a credible long-term owner remain essential.