← /editions
#14

Claude just went into the lab

// Anthropic launched Claude Science in beta, a multi-agent workbench that delegates from a coordinating agent to domain specialists in genomics, proteomics, and cheminformatics, with a reviewer agent checking citations and calculations and reproducible provenance on every figure — a production example of harness-over-model design for high-stakes work.

$ today_in_10_seconds

Anthropic launched Claude Science in beta — a multi-agent workbench that drops a coordinating agent, domain specialists, and a citation-checking reviewer agent straight into genomics, proteomics, and cheminformatics workflows. The design bet is the one this newsletter keeps circling: real autonomy in high-stakes domains needs specialist sub-agents, a review gate, and reproducibility baked in. It's agentic AI aimed squarely at the lab bench.

$ lead_story

Anthropic launches Claude Science, a multi-agent workbench for the lab

Anthropic released Claude Science in beta, a desktop app (macOS/Linux) that puts a generalist coordinating agent in front of 60+ curated skills and connectors pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics. The architecture is textbook hierarchical multi-agent: the coordinating agent breaks a plain-language request into subtasks and delegates to domain-specialized sub-agents, each with context about its inputs and the databases it connects to — and, crucially, a separate reviewer agent checks every citation and calculation and fixes errors as it goes. The part builders should study is provenance: every figure ships with its exact code, environment, description, and full message history, and compute runs locally, on HPC over SSH, or bursts to Modal from one GPU to hundreds. It's not a new model — it runs on existing Claude models including Opus 4.8 — it's a harness.

// why_it_matters: This is the clearest production example yet of the pattern that actually makes agents trustworthy in high-stakes work: specialist sub-agents, a dedicated reviewer, and reproducibility as a first-class output.

read more →

$ also_worth_attention

  • Anthropic pairs the launch with up to $30K in compute grants for scientists

    Alongside the app, Anthropic's AI for Science program is offering compute grants of up to $30,000 for researchers with active graduate or postdoctoral projects, with a July 15, 2026 deadline and Modal adding $2,000 in compute for select projects. The initial focus is cross-disciplinary biology and biomedical work, with selected projects running September 1 to December 1. The tell for founders: distribution strategy for agentic tools increasingly means subsidizing the compute that makes long-horizon runs viable, not just handing over software.

    link →
  • The BioNeMo integration is where the speed comes from

    Claude Science leans on NVIDIA's BioNeMo Agent Toolkit, exposing GPU-accelerated models as callable skills — Evo 2 for genomic sequence analysis, Boltz-2 for structure prediction, and OpenFold3 for protein folding. The number that lands: NVIDIA's RAPIDS-singlecell toolkit compresses a 1.3-million-cell preprocessing and clustering workflow from 52 minutes to 25 seconds. The lesson that generalizes beyond biology — an agent is only as useful as the specialized, high-performance tools it can call, and grounding it in a real toolchain beats raw model horsepower.

    link →
  • The "AI co-scientist" race is getting crowded

    Claude Science lands in a fast-filling field of autonomous-research systems — Google DeepMind's AI Co-Scientist for hypothesis generation, Sakana's AI Scientist, and a wave of 2026 arXiv work on multi-agent 'AI researchers.' The recurring critique in that literature is worth carrying into any eval: many systems optimize for producing papers or benchmark scores on tasks that finish in minutes, not the long, messy, verifiable work real discovery demands. Anthropic's bet on reproducibility and a reviewer agent is a direct answer to that critique — whether it holds up under independent testing is the open question.

    link →

$ bigger_picture

The trend behind the headlines

Claude Science is the same lesson this newsletter keeps landing on, now shipped as a product: in high-stakes domains, the model is table-stakes and the harness is the product. Look at what Anthropic actually built around its existing models — specialist sub-agents that know their tools, a reviewer agent as a hard check on citations and math, provenance attached to every figure, and flexible compute from laptop to HPC. That's a direct descendant of the security-and-grounding themes from recent editions: human-in-the-loop review, auditable history, and controls that catch a wrong answer before it ships. It also closes a loop we've traced from Tecan's lab agents (Edition #9) through the BioNeMo toolkit — the same GPU-accelerated science skills now sit under a general-purpose reasoning agent. The frontier isn't a smarter model; it's the trustworthy scaffolding that lets one act in a domain where being confidently wrong has real consequences.

$ tool_of_the_day

NVIDIA BioNeMo Agent Toolkit

The engine room beneath today's lead, and a returning guest in this newsletter for good reason. It packages GPU-accelerated life-sciences models (Evo 2, Boltz-2, OpenFold3) as skills an agent can call, which is exactly why Claude Science can go from plain-language request to a rendered 3D protein structure with reproducible code. Even outside biology, it's the reference design worth internalizing: ground your agent in a curated, high-performance toolset rather than expecting the base model to know everything.

explore NVIDIA BioNeMo Agent Toolkit →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime