AI agents transform scientific computing with new capabilities and challenges

Scientific software, frequently built as companion code to research papers by small academic teams, has long struggled with maintenance and deterioration as data generation outpaces engineering support. AI agents are reshaping this landscape by reducing engineering costs and accelerating prototyping cycles. An exploratory report examined eight agent-assisted scientific computing projects, primarily in life sciences, using tools such as Codex and Claude Code. Projects ranged from routine maintenance to major undertakings like language migrations and GPU-native redesigns. Agents significantly accelerated software development and maintenance tasks, enabling small teams to accomplish work that previously required specialized engineering expertise. The Python library cyvcf2 exemplified this shift, modernizing its build and packaging system entirely. However, contributors identified a persistent challenge: unclear long-term responsibility and stewardship of resulting tools remain unresolved. A consistent pattern emerged across projects: researchers transitioned from direct implementation roles toward verification and orchestration, defining objectives, measuring correctness, and determining release readiness. Agents, despite their speed, cannot reliably validate whether output is scientifically sound. Humans must therefore develop robust validation approaches using external references or measurable targets such as exact output agreement or parity with existing tools. Projects typically proceeded through iterative, feedback-driven stages rather than single-shot attempts. Teams broke broad goals into smaller increments, using intermediate benchmarks and test systems to evaluate and refine agent work. While agents produced initial implementations quickly, resolving edge cases and subtle numerical discrepancies demanded substantially more effort. These case studies demonstrate that agents are easing engineering constraints, allowing researchers to focus more on scientific questions while maintaining control over quality and long-term stewardship of their tools.