Anthropic wrote in a science blog that AI agents excel at coding tasks but stumble in biological retrieval, and the root cause is not the models' intelligence but the data and APIs. Biological databases were designed for human web browsing rather than direct agent access, causing frontier models to produce inconsistent retrievals when asked to fetch virus sequences.
What happened
In Anthropic's experiments, the same prompt used to retrieve virus data produced wildly different results: on one Ebola task the models returned 266 records, then 106 records, and later just 5 records from the identical prompt. Introducing a tool called gget virus, which provides deterministic retrieval, raised every evaluated model above 90% performance and erased the performance gap between inexpensive and expensive models.
Why this matters
The industry tendency is to await the next, more capable model. Anthropic's findings point in the opposite direction: fixing the plumbing of data access made model choice largely irrelevant. With deterministic, agent-friendly retrieval, a cheaper model plus the right tool matched flagship models. The blocker, therefore, was not intelligence but that the world’s data infrastructure is geared toward human interactions, not automated agents.
Parallels and implications
Andrej Karpathy has voiced a similar complaint regarding software: writing code can be straightforward, while the real time sink is navigating setup screens and interfaces. In biology the stakes are higher — poorly accessible data can lead not just to lost developer time but to missed diagnoses or other consequential failures.
Conclusion
Anthropic suggests that the next major advance in applied AI may come less from larger models and more from building reliable, auditable, agent-oriented data and API plumbing — the unglamorous infrastructure that enables consistent, trustworthy automation.
