Anthropic Announces gget virus: Deterministic Retrieval Layer Boosts AI Agent Accuracy for Viral Sequence Data

TL;DR

Anthropic released gget virus, a deterministic retrieval layer for NCBI Virus that lifts AI agent accuracy on viral sequence retrieval tasks to >99%, demonstrating that reliable data‑access tools are essential for trustworthy biological AI agents.

The core problem: brittle biological data infrastructure

Biological databases such as NCBI Virus were built for human users, not for autonomous agents.
Heterogeneous file formats, scattered APIs, and undocumented filtering logic create “click‑tax” workflows that are easy for humans but error‑prone for AI agents.
Even state‑of‑the‑art scientific agents (Claude Sonnet 4, Claude Opus 4.7, Biomni OSS, Edison Analysis, GPT‑5.2‑pro, GPT‑5.5) achieved mean accuracies of 16.9 %–91.3 % on realistic viral‑sequence queries, with large run‑to‑run variability.
Small retrieval errors (e.g., wrong genome build, missing records) can drastically alter downstream conclusions such as phylogenetic TMRCA estimates or therapeutic epitope analyses.

Case study: VirBench benchmark on NCBI Virus

VirBench comprises 120 queries across 40 pathogens, each with manually verified ground‑truth counts.
Example query: retrieve Orthoebolavirus zairense sequences from human hosts in Africa (01/01/2014–06/20/2014) with length ≥ 15,200 bp and ≤ 1,900 N’s.
Agents without deterministic support returned wildly inconsistent counts (e.g., Claude Sonnet 4 gave 106, 15, and 5 sequences across three identical runs).
Resulting phylogenetic trees showed TMRCA estimates ranging from 1922 to 2014, illustrating how retrieval gaps corrupt scientific inference.

"The code was the easiest part! Most of the work was in the browser, clicking things." – Andrej Karpathy, cited to illustrate the broader friction between agents and human‑centric tools.

Introducing gget virus

Developed with NCBI collaborators to reproduce the NCBI Virus web‑interface behavior programmatically.
Coordinates multiple underlying resources (REST, Datasets, E‑utilities) and applies filters locally when APIs lack the needed semantics.
Handles large result sets via batching, resolves identifier mismatches, and preserves metadata (e.g., GenBank fields) in a standardized output.
Emits detailed logs that make the retrieval process auditable and reproducible.

Impact on agent performance

Agent	Mean accuracy without gget virus	Mean accuracy with gget virus
Claude Sonnet 4	91.3 % (variable)	99.5 %
Claude Opus 4.7	88.7 %	99.6 %
Biomni OSS	84.2 %	99.3 %
Edison Analysis	78.5 %	99.2 %
GPT‑5.2‑pro	85.1 %	99.4 %
GPT‑5.5	89.9 %	99.7 %

Adding gget virus eliminated most run‑to‑run variability; the performance gap between models narrowed dramatically.
The result shows that deterministic retrieval, not model size, is the limiting factor for reliable biological data pipelines.

Why deterministic layers matter

Biological workflows demand near‑perfect recall; a single missing sequence can mislead outbreak timing or therapeutic efficacy assessments.
Deterministic tools provide a reproducible “highway tunnel” beneath the fragmented “pedestrian streets” of legacy bioinformatics portals.
They enable cheaper models to achieve expert‑level reliability, lowering cost and widening access to AI‑augmented research.

Broader implications for AI‑driven science

The mismatch between agent capabilities and existing infrastructure mirrors challenges in other domains (e.g., web development, as highlighted by Karpathy’s lecture).
As models improve, the need for harnesses may diminish, but deterministic interfaces will remain valuable for auditability, speed, and cost‑effectiveness.
Future database design should prioritize agent‑friendly APIs, standardized identifiers, and version‑controlled metadata to avoid the “click tax” altogether.

Next steps and open challenges

Extend the deterministic‑layer approach to other domains (e.g., protein structure repositories, metabolomics databases).
Develop community standards for “agent‑ready” data services, including explicit schema versioning and machine‑actionable documentation.
Monitor how rapidly improving models interact with deterministic tools; assess whether tool‑model co‑evolution can further reduce latency and cost.

Acknowledgements: The authors thank Xander Balwit, Ethan Dyer, Stuart Ritchie, Rebecca Hiscott, Alyssa Morrow, Keir Bradwell, Eric Kauderer‑Abrams, Jonah Cool, Andrej Karpathy, Patrick Varilly, Cesar Arze, Blake Lash, Philine Guckelberger, Nisha Gopal, Elliot Hershberg, Pardis Sabeti, Jonathan Feldman, Sarah Gurev, Gage Moreno, Ferdous Nasri, and Krithik Ramesh for feedback and contributions.