AI agents put science’s validation bottleneck in focus

AI agents are moving beyond chatbot-style answers into goal-directed research systems that can plan, call tools, run subagents, search literature, write code and correct errors. Google DeepMind’s Co-Scientist matched an unpublished Imperial College London hypothesis about how superbugs spread antibiotic resistance, while researchers have used similar systems to generate drug-repurposing ideas and candidate algorithms.

The shift is being driven by stronger frontier models, inference-time reasoning, scaffolding that gives models memory and tool access, and customizable “skills” that capture parts of a lab’s workflow. Agents are already helping scientists sift papers, query databases, coordinate analyses, curate data and build software pipelines. They can also explore large solution spaces through systems such as AlphaEvolve, which has supported chip design, mathematics work and genomics analysis.

The main bottleneck is validation. AI agents can make hypotheses and candidate solutions abundant, but testing them still depends on experiments, peer review and institutional capacity. Mathematics and computer science benefit from formal verification, yet many fields still require wet labs, physical measurements or slow biological processes. Policy priorities include wider access to agents, agent-ready datasets, more experimental infrastructure, automated labs and updated peer review practices that let reviewers use agents transparently.