Scientific Computing in the Age of Agentic AI
A field report across eight scientific-software projects finds that coding agents can remove substantial engineering friction while shifting the human bottleneck toward specification, validation, and long-term stewardship.
Which acceptance tests best predict scientific validity when an agent can satisfy software tests while still implementing the wrong scientific assumption?
This field report covers eight agent-assisted scientific-computing projects, primarily in the life sciences. Five used Codex alone; three combined Codex and Claude Code. The work ranged from packaging and maintenance to language migration, performance optimization, and GPU-oriented redesign.
Project teams report meaningful acceleration, especially where a small scientific group lacked dedicated software-engineering capacity. The report is exploratory rather than a controlled productivity study, and the contributors identify a persistent limit: agents could implement well-scoped requests but could not reliably judge scientific correctness on their own.
The validation pattern
The strongest projects used explicit acceptance targets: exact agreement with an existing tool, known answers from simulated data, statistical properties, or performance parity under controlled tests. Human researchers moved from direct implementation toward specification, orchestration, review, and stewardship.
ASI relevance
Scientific code is leverage for the entire research process. Agents that modernize and extend it can make more experiments affordable, but software correctness and scientific validity are not the same. The next systems need evaluators that can detect a plausible implementation of the wrong assumption.