Coding agents accelerate scientific software development

The report tracks eight projects across genomics, immunology, and statistics, utilizing OpenAI’s Codex and Anthropic’s Claude Code. While the results show significant performance gains—such as a 31 percent runtime reduction for the DNA-sequencing tool HI.SIM and a 60-fold speed increase for the quality-control suite RustQC—contributors emphasize that agents serve as accelerators rather than autonomous replacements for expert oversight.

Successful implementations relied on a clear division of labor. Agents handled repetitive coding tasks, such as porting TensorFlow backends to PyTorch for MHCflurry or optimizing statistical models in Rust, but human contributors remained essential for defining benchmarks and validating scientific accuracy. The shift from manual coding to steering agent-driven drafts allows researchers to attempt complex rebuilds of abandoned tools, like the STAR aligner, that were previously deemed too time-consuming for small teams.

Despite the efficiency, experts warn that low-cost code generation introduces new risks. Rapidly rebuilding and modifying software can lead to fragmented toolsets where different labs end up with incompatible versions of the same instrument. As contributors noted, the primary constraint has shifted from writing code to rigorous verification and long-term stewardship. The findings suggest that before deploying agents to modernize research infrastructure, teams must establish clear ownership and testing protocols to ensure that speed does not come at the expense of scientific consistency.