A new field report examining eight agent-assisted scientific computing projects-primarily in the life sciences-finds that AI coding agents are reshaping how research software gets built and maintained. Five projects used Codex alone, three combined Codex with Claude Code, and the results point to a practical shift: researchers spend less time on implementation and more time directing the scientific work.
The report, which draws on case studies written by the teams behind each project, covers work ranging from routine maintenance and targeted optimization to large-scale language migrations and GPU-native redesigns. In several cases, small teams took on work that would have required far more time or specialized engineering support without agent assistance.
The changing role of the researcher
Across all eight projects, contributors described a consistent pattern. Researchers moved from writing code to specifying what to build, defining how to measure correctness, and deciding when a project was ready to ship. The agents handled tedious implementation tasks, but humans stayed in control of the scientific direction and quality bar. This emerging model keeps verification and orchestration firmly in the researcher's hands while AI Agents & Automation provide velocity on the engineering side.
"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," said Brent Pedersen, a contributor to one of the projects.
Modernizing genomic data tools
One case study focused on cyvcf2, a Python library for reading and writing genomic variant files. GPT-5.5 replaced the library's legacy build and packaging system with a modern, unified process designed to make the library easier to install, test, and release. Other projects included performance-based refactoring and rewrites that reduced computing demands or brought abandoned tools under new community stewardship.
Verification becomes the bottleneck
While agents handled specific, well-scoped requests effectively, they could not reliably judge whether their work was scientifically valid. Agents often expressed confidence even when their output contained clear errors. Human reviewers had to find reliable ways to validate results-using external references, exact output agreement with existing tools, or answers established in advance with simulated data.
The projects generally proceeded in feedback-driven iterations rather than one-shot attempts. Contributors broke broad goals into smaller changes and used intermediate benchmarks to evaluate the agents' work. Agents produced initial implementations quickly, but resolving edge cases and subtle numerical differences took much longer. The "last mile" of an implementation consistently demanded the most effort.
Long-term stewardship remains essential
The maintenance gap in research software has long slowed iteration and limited reproducibility. Published studies have found that research code often fails to install properly in a fresh computing setup or run as documented, forcing researchers to spend substantial time on configuration and debugging. Lower implementation costs through agents could make it easier to produce many similar rewrites, fragmenting users and spreading expert attention thin.
Mature scientific software carries undocumented conventions, compatibility requirements, and user trust that translating source code alone cannot reproduce. Changes to some projects, such as MHCflurry and cyvcf2, were incorporated into their original upstream repositories. Another project, rustar-aligner, moved under new community stewardship after the original project was abandoned. Where coordination with existing maintainers is possible, the report recommends starting it early. When a separate implementation is necessary, it needs a clear owner and a credible maintenance plan.
Why this matters for science and research professionals
The deeper change documented in these case studies is not simply that researchers can produce more software. Coding agents are shifting where researchers direct their effort-away from implementation details and toward defining goals, validating outputs, and planning for long-term stewardship. For AI for Science & Research workflows, the practical takeaway is that verification skills and maintenance planning are becoming as important as coding ability. Researchers who build reliable validation methods and establish clear ownership early will get more durable results from agent-assisted development.
Your membership also unlocks:
