Paving the way for agents in biology \ Anthropic<br>Try Claude

Science<br>Paving the way for agents in biology<br>Jun 8, 2026

Written by Laura Luebbert. Based on research by Ferdous Nasri, Sarah Gurev, Patrick Varilly, Krithik Ramesh, Nuala A. O’Leary, Jonah Cool, Bernhard Y. Renard, Pardis Sabeti, and Laura Luebbert.

In this post, Laura Luebbert argues that we need to make biological data infrastructure more agent-friendly. As a case study, she and her team tasked scientific research agents (Claude, Biomni, Edison Analysis, GPT) to retrieve the sequence data from NCBI Virus, a database virologists use for tasks such as surveillance and diagnostic assay development. Even the strongest models did not consistently achieve the level of accuracy required for reliable dataset construction. But accuracy rose to nearly 100% once she and her team added gget virus, a deterministic retrieval layer. The broader lesson for scientific agents is that deterministic retrieval tools are (currently) crucial to making agent workflows more reliable, and biological databases will need to be designed with agents in mind as scaled users.

Using AI agents to navigate biological data infrastructure is like driving through an old city that was designed before cars: the infrastructure may be beautiful and even thoughtful, but it’s full of narrow, winding streets that are difficult for modern vehicles to navigate (idiosyncratic file formats, scattered databases, and one-off retrieval scripts).1 You can retrofit the city with traffic signs, parking lots, and the occasional widened road, but the basic layout remains hard to navigate because it was designed for a different mode of conveyance. Software infrastructure, by contrast, was basically made for the needs of cars (agents): paved roads, clear lanes, standardized signals, and systems designed for fast travel from start to finish (version control, well-documented APIs, and package managers).<br>As a result, coding agents have advanced much more quickly than biological agents. Software commonly provides structured digital workflows and reliable interfaces, whereas the computational biology infrastructure needed for data retrieval and validation is often brittle, heterogeneous, and process-dependent. The tools with which we navigate them are necessarily bespoke and tuned to defined domains or hypotheses. Moreover, software provides testable outputs that can be quickly compiled and validated (e.g., resolving a GitHub issue by generating a patch that passes the project’s tests), whereas biology offers few simple and verifiable yet meaningful rewards.<br>Thus, the bottleneck for biological agents is not only reasoning but the absence of widespread deterministic execution layers for querying biological data. A scientist can express their intent (e.g., find all human kinases with this domain and pull their structures), but agents often lack a dependable way to access the databases containing the information they need.<br>In biological and scientific workflows, even small errors can have severe consequences. Retrieving coordinates from the wrong genome build, for example, can invalidate the downstream biological interpretation. So can mixing RefSeq and GenBank records without intending to, treating partial genomes as complete genomes, confusing segment names in segmented viruses, or missing relevant records because of inconsistent metadata fields. The beauty and challenge of research is that the details are often of critical importance.<br>Like driving through an Italian hill town, it does not matter how powerful the car is if the streets are too narrow, the turns too sharp, and the route depends on local knowledge. If we want agents to help with scientific discovery, from outbreak response to drug design to biological modeling, we need to build biological data infrastructure that they can navigate as reliably as humans do.<br>What Karpathy’s lecture about web development tells us about doing biology with AI agents

This mismatch between agent needs and human-built tools is not unique to biology. The same friction emerges wherever agents are inserted into environments designed solely for human use.<br>A few months ago, Andrej Karpathy gave a talk about software in the era of AI and ended up griping about something that sounded all too familiar. He had vibe-coded a small web app, but when he tried to make it real (authentication, payments, deployment), he lost a week clicking around in browser dashboards.<br>As he summarized, “The code was the easiest part! Most of the work was in the browser, clicking things.” Documentation kept telling him to “go to this URL, click on this dropdown.” His conclusion was that nobody should have to do this. Instead, we must build for agents.<br>Karpathy had experienced something new within the world of software agents that biology researchers have been struggling with for a long time: the pain of trying to make intelligent systems operate in environments built around heterogeneous...