Laying the Groundwork for AI Agents in Biology

Biology data infrastructure is like an old city that was never designed for cars to drive through. File formats are idiosyncratic and databases are fragmented, creating a structure that is extremely difficult for AI agents to navigate.

Software infrastructure, by contrast, has standardized tools well-suited for agents to use, such as APIs, version control, and package managers. Because of this, coding agents are advancing rapidly, while biology agents are progressing more slowly due to the complexity of data retrieval and validation processes.

When the research team ran experiments extracting data from NCBI Virus using state-of-the-art models such as Claude, Biomni, and GPT, accuracy came out low—not because of the models' reasoning ability, but because of instability in the data access method. However, once a deterministic retrieval layer called gget virus was added, accuracy reached nearly 100%.

In biology data, small errors (such as using the wrong genome build, or mixing up RefSeq and GenBank) matter so much for precision that they can invalidate an entire study. Therefore, in order for agents to assist in scientific discovery, it is essential to build agent-friendly infrastructure that includes a deterministic execution layer.