Paving the way for agents in biology
Anthropic argues biology needs more agent-friendly infrastructure. In a virology benchmark, deterministic retrieval pushed accuracy from inconsistent results to nearly 100%.
Intelligence analysis by GPT-5.4 Mini
The essay says AI agents are held back in biology not just by model reasoning, but by messy databases, inconsistent metadata, and browser-heavy workflows. A deterministic retrieval layer, gget virus, made viral sequence fetching far more reliable.
The article says AI helpers are like smart kids trying to find books in a library where the shelves are messy and the labels are confusing. If the library is fixed with clear signs and simple rules, the helpers can find the right books much more often.
Analysis
The core argument
Anthropic’s Laura Luebbert argues that biological data systems are not built for agents. Unlike software, which usually has APIs, version control, and predictable workflows, biology data often lives across scattered databases, inconsistent identifiers, and manual web interfaces. That makes it hard for agents to retrieve, validate, and repeat scientific work reliably.
The VirBench case study
The post uses viral sequence retrieval from NCBI Virus as a test case. The team asked several scientific research agents, including Claude, Biomni OSS, Edison Analysis, and GPT, to complete realistic retrieval tasks. Even the strongest models did not reliably reach the level of accuracy needed for dataset construction. The article says mean accuracy ranged from 16.9% to 91.3%, and the same model could return very different answers across repeated runs.
That unreliability matters because sequence retrieval is often the first step in a larger scientific workflow. The essay shows how missing or incorrect sequences can distort downstream analyses, including phylogenetic trees and estimated outbreak timing. In one example, different retrieval runs produced very different interpretations of the 2014 Ebolavirus outbreak.
Why deterministic tools help
When the team added gget virus, a deterministic retrieval layer built with researchers at NCBI, accuracy rose above 90% for all agents and reached 99.7% for GPT-5.5. The article frames this as evidence that scientific agents need boring, reliable infrastructure underneath them. The point is not that models are useless, but that they need machine-actionable tools that can handle batching, filtering, metadata quirks, and reproducible outputs.
The broader lesson
The essay extends the lesson beyond virology: if AI agents are expected to help with outbreak response, drug design, or biological modeling, the databases they use have to be designed with agents in mind. Otherwise, even capable models will keep producing answers that look plausible but are wrong.
Key points
- Anthropic says biology needs infrastructure designed for agents, not just stronger models.
- A benchmark called VirBench tested viral sequence retrieval from NCBI Virus.
- Without deterministic support, the agents were inconsistent and often inaccurate.
- Adding gget virus raised accuracy above 90% for all agents and near 100% for the best result.
- The article argues that reliable science depends on reliable data retrieval, not just reasoning.
If more biological databases add deterministic, agent-friendly layers, scientific agents could become much more reliable for tasks like sequence retrieval and data curation. That could make outbreak analysis, assay design, and biological modeling faster and more repeatable, even with cheaper models.
If the infrastructure stays messy, agents may keep producing plausible but wrong scientific datasets, which can distort conclusions about outbreaks or therapeutics. Even with helper tools, the article notes that the systems can still be too expensive, too slow, or too hard to audit for routine scientific work.
