Why Biology AI Agents Stall: The Data Infrastructure Bottleneck, Not Model Size

Anthropic’s new research blog argues that the slow progress of AI agents in biology is not due to weak large‑model reasoning but to the outdated and heterogeneous biological data infrastructure that agents must navigate.

Data Infrastructure vs. Modern Software

The author likens trying to drive an AI agent through current biological databases to navigating a narrow, winding old city built before cars existed—full of proprietary file formats, scattered databases, and one‑off scripts—whereas software development enjoys well‑structured version control, clear APIs, and package managers that act like paved roads for agents.

Consequences of Fragile Workflows

Because biological tools lack a deterministic execution layer, agents cannot reliably retrieve the needed information. Small mistakes—such as using the wrong genome version or mixing RefSeq with GenBank records—can invalidate downstream analyses.

Karpathy’s “Click Tax” Analogy

The article draws a parallel to Andrej Karpathy’s complaint about web‑development “click tax,” where most effort is spent clicking through browsers rather than coding. The same friction appears in virus research, where scientists must manually apply dozens of filters on the NCBI Virus web interface.

Case Study: Virus Data Retrieval

Using the 2026 Bundibugyo Ebola outbreak as an example, the author shows that answering urgent questions (variant comparison, diagnostic coverage, therapeutic efficacy) requires precise NCBI Virus queries, yet the manual filtering process is error‑prone and time‑consuming.

VirBench Benchmark

To quantify the gap, the team built VirBench, a benchmark of 120 realistic virus‑sequence queries covering 40 pathogens. Model performance ranged from 16.9 % to 91.3 % accuracy, and repeated runs of the same prompt produced wildly different results—for example, Claude Sonnet 4 returned 106, 15, and 5 sequences for the same Ebola query, despite identical prompts.

Impact on Downstream Analysis

These inconsistencies affect phylogenetic trees and TMRCA estimates; the same model sometimes inferred a common ancestor in 2014, while a faulty run pushed it back to 1922.

gget virus: A Deterministic Retrieval Layer

In response, the researchers collaborated with NCBI to create gget virus, a tool that orchestrates REST, Datasets, and E‑utilities APIs, resolves complex filters locally, handles pagination, and outputs standardized, logged results. This deterministic layer lets agents verify and reproduce their answers.

Performance Gains

After integrating gget virus, all tested agents achieved >90 % accuracy, with GPT‑5.5 reaching 99.7 %, and run‑to‑run variance virtually disappeared, narrowing the gap between model capabilities.

Broader Implications

The author concludes that while future agents may eventually bypass such tools, reliable, reproducible data infrastructure remains essential for scientific discovery, and building “context engines” for biological data is a critical direction for AI‑driven research.