Speaker
Description
Research Software Engineers frequently inherit scientific codebases they did not write, in unfamiliar domains, and on schedules that limit deep familiarization. Large Language Model–based coding agents, such as Codex, can accelerate this transition, but their value depends on how they are integrated with scientific verification and software-development practices. This presentation examines the use of agentic AI to assess, extend, and rewrite two scientific Python repositories supporting Europa Clipper mission planning.
The first repository was an approximately 60,000-line implementation of the Hapke bidirectional reflectance model developed by a domain scientist. An AI-assisted audit traced equations through the implementation, compared the code with published model descriptions, and generated targeted questions for a subject-matter expert. During the initial audit, the agent identified an algebraic error that was independently verified against the scientific literature. The repository otherwise remained largely intact. Here, the agent’s principal value was accelerated code comprehension and preparation for expert review rather than autonomous code generation.
The second repository, originally a data-ingestion wrapper, required a near-complete rewrite to support synthetic spectral generation, photometric and geometric calculations, and observation-planning workflows. Coding agents accelerated bug identification, implementation, testing, and refactoring. However, rapid AI-assisted development also introduced redundant utilities, inconsistent abstractions, and growing architectural complexity. Structured practices drawn from Git.Ship.Done. were subsequently used to organize work into documented requirements, plans, and executable tasks, improving continuity and making agent behavior easier to review.
These case studies show that agentic AI can lower the barrier to working with unfamiliar scientific software, improve interactions with domain experts, and accelerate delivery of complex capabilities. They also demonstrate that agents can amplify weak architectural decisions and cannot replace scientific validation, code review, or deliberate software design. The presentation will discuss effective workflows, verification practices, and the continuing challenge of building maintainable and reproducible agent-assisted scientific software.