AI/TLDR

Anthropic · 2026-06-08 · notable

Anthropic Research: 'Paving the Way for Agents in Biology' — Adding a Deterministic gget virus Layer Lifts NCBI Virus Retrieval Accuracy From a Floor of 16.9% on Claude Opus 4.7 to Above 90% Across Sonnet 4, Opus 4.7, GPT-5.2-pro, GPT-5.5, Biomni OSS, and Edison Analysis, Peaking at 99.7%

Anthropic and pachterlab benchmark six science agents on NCBI Virus retrieval and show all of them fail without help. Plugging in a deterministic context engine (gget virus) lifts every agent above 90% accuracy, peaking at 99.7%.

Anthropic Research illustration for 'Paving the way for agents in biology' on an olive background

A Bundibugyo Ebola outbreak motivates a hard look at why AI agents still can't reliably pull viral sequences out of NCBI.

Key specs

Baseline accuracy16.9%-91.3%
With gget virus>90% all agents
Peak accuracy99.7%

What is it?

An Anthropic Research essay co-authored with the Pachter Lab and outbreak responders that benchmarks six science agents — Claude Sonnet 4, Claude Opus 4.7, GPT-5.2-pro, GPT-5.5, Biomni OSS, and Edison Analysis — on viral sequence retrieval tasks against the public NCBI Virus database. It then ships gget virus, a deterministic Python tool that wraps NCBI's retrieval surface in a stable, agent-friendly interface.

How does it work?

The team gave each agent the same viral-retrieval task list and let it write code against NCBI's existing APIs. Bare-agent accuracy ranged from 16.9% (Claude Opus 4.7) to 91.3%, with high run-to-run variance. Re-running the same tasks with gget virus available collapsed that variance and pushed every agent above 90%, peaking at 99.7%. The paper argues bioinformatics needs more 'context engines' — small, deterministic interfaces that agents can call instead of reasoning over scattered, idiosyncratic data formats.

Why does it matter?

Science-agent demos almost always assume a smarter model is the only missing piece. This benchmark argues the bottleneck is infrastructure: even frontier models like Opus 4.7 and GPT-5.5 fail on a 'simple' viral retrieval task without a context engine. The Bundibugyo outbreak framing — the DRC has logged more than 1,000 suspected cases and 200 deaths since the May 14 INRB sequencing — gives the argument urgency. gget virus is shipping now via the existing pachterlab/gget package.

Who is it for?

computational biologists, bioinformatics tool builders, AI-for-science researchers

Try it

pip install gget && python -c "import gget; gget.virus('Bundibugyo virus', host='Homo sapiens', nuc_completeness='complete')"

Sources · 3 outlets

Tags

  • anthropic
  • claude-opus-4-7
  • biology
  • agents
  • ncbi-virus
  • gget
  • scientific-tools
  • context-engineering
  • research

← All releases · Learn AI