This Anthropic Science post argues that the main obstacle to AI agents doing real biology work is not the reasoning ability of the underlying models but the state of biological data infrastructure, which was built for humans clicking through web pages rather than for agents querying at scale. The authors compare it to driving a modern car through an old city designed before cars existed.
To test this, the team built a benchmark (VirBench) around retrieving viral sequence data from NCBI Virus, a database virologists use for surveillance and diagnostic work. Run directly, frontier agents including Claude Sonnet 4, Claude Opus 4.7, GPT-5.2-pro, and GPT-5.55 achieved mean accuracies ranging from just 16.9 percent to 91.3 percent on identical repeated queries, far too unreliable for building real datasets.
The fix was a deterministic retrieval tool called gget virus, built with NCBI, that coordinates NCBI’s REST, Datasets, and E-utilities APIs and returns standardized output. With the tool, every agent crossed roughly 90 percent accuracy, peaking at 99.7 percent for GPT-5.5. The takeaway is that deterministic execution layers, not just smarter models, are what make agentic science reliable, and biological databases should be designed with agents as first-class users.