gget virus, Anthropic's deterministic layer to make AI agents reliable in biology

Anthropic and NCBI developed gget virus, a deterministic layer raising AI agent accuracy to 99.7% for viral sequence retrieval on the VirBench benchmark.

Anthropic publishes a study where Laura Luebbert and her team argue that biological data infrastructure needs to be rethought for AI agents. The initial observation: confronted with a seemingly basic task, retrieving viral sequences from the NCBI Virus database, the best current scientific agents (Claude, Biomni OSS, Edison Analysis, GPT) are not reliable. On their VirBench benchmark, which comprises 120 realistic queries covering 40 pathogens with manually verified counts, average accuracies range from 16.9% to 91.3%, and the same model often returns different answers to an identical query posed three times.

The cause lies less in reasoning than in the absence of a deterministic execution layer: a large part of NCBI Virus's filtering logic exists only within its web interface, making it a nightmare to automate, exactly the "click tax" that Andrej Karpathy described for software development. The consequences are tangible: for Ebolavirus sequences related to the ongoing Bundibugyo virus epidemic in the DRC, incomplete datasets shifted the estimated origin date of the epidemic or skewed the analysis of therapeutic antibody targets.

To remedy this, the team developed gget virus in collaboration with NCBI, a layer that orchestrates underlying APIs (REST, Datasets, E-utilities), manages large-volume batching, and returns standardized and traceable outputs. Once this tool was provided to the agents, accuracy exceeded 90% for all, culminating at 99.7% with GPT-5.5, and variability between executions disappeared, making the choice of model almost indifferent. The authors nevertheless qualify this: even if models eventually learn to navigate these portals independently, delegating the task to an agent each time would remain too costly, too slow, and too difficult to audit for routine scientific use.