Scientific computing, a cornerstone of research, has been hampered by slow, fragile software workflows. AI agents are beginning to change this by lowering engineering costs and automating tedious tasks, allowing researchers to prototype faster and maintain software more effectively. An exploratory report details eight agent-assisted projects, primarily in life sciences, using Codex and Claude Code. These projects, ranging from library modernization to large-scale migrations, show agents significantly accelerating development and maintenance. Contributors report a shift in their role from implementation to verification and orchestration, with researchers defining goals and judging correctness while agents provide velocity.

Key themes include the necessity of human judgment for validating AI output, as agents can generate errors with confidence. Strong validation approaches use external references or measurable targets. Projects proceeded iteratively, with agents producing initial implementations quickly, but resolving edge cases required significant human effort. The "last mile" of implementation proved most challenging.

Long-term stewardship remains crucial. While agents ease implementation burdens, ensuring the reliability and maintenance of scientific software requires clear ownership and credible plans. The case studies suggest agents enable researchers to focus more on scientific direction and validation, accelerating iteration and advancing fields.