Claude Science: reproducibility as a first-class artifact

Most "AI for science" announcements describe a future. The one Anthropic published on June 30 describes a tool. Claude Science is an app — in beta on macOS and Linux for Pro, Max, Team, and Enterprise — that tries to collapse the scattered apparatus of computational research into a single agentic environment. I've spent enough time in both physics and production agent systems to be wary of the genre, so I read the launch looking for the engineering decisions rather than the adjectives. A few are worth your attention.

What actually shipped

The pitch is mundane in the best way. A working scientist juggles dozens of databases with incompatible schemas, file formats that each need their own viewer, and a rotating cast of tools — PubMed, Jupyter, R, a cluster terminal. Claude Science puts a generalist coordinating agent in front of all of it, with over 60 curated skills and connectors pre-wired for genomics, single-cell, proteomics, structural biology, and cheminformatics, backed by more than 60 scientific databases (UniProt, PDB, Ensembl, ChEMBL, GEO, and the like). You ask in plain language; specialist sub-agents query and synthesize across the sources. It connects to NVIDIA's BioNeMo toolkit, so models like Evo 2, Boltz-2, and OpenFold3 are reachable from the same conversation. And it renders what scientists actually look at — 3D protein structures, genome browser tracks, chemical structures — natively rather than as a wall of text.

The shape of the system: one coordinating agent, specialist sub-agents over curated data and compute, a reviewer in the loop, and a reproducible artifact at the end.

That's a competent integration story. It is not, by itself, the reason to care.

Reproducibility as a first-class artifact

Here is the design choice I'd underline. When Claude Science produces a figure, it bundles the exact code and environment that generated it, a plain-language description of how it was made, and the full message history. Every result is traced back to the code that produced it.

I spent years in physics watching reproducibility erode at exactly this seam — the gap between the figure in the paper and the unrecoverable sequence of scripts, manual tweaks, and forgotten parameters behind it. Six months later nobody, including the author, can rebuild it. Making provenance a structural part of every output — not an optional methods section written from memory — is the right place to attack the problem. You can ask it to change an axis to log scale or drop the gridlines, and it edits its own code rather than hand-painting the image, which keeps the artifact and its provenance in sync. That detail matters more than it looks: an audit trail is only worth something if it can't drift from the result.

Keep the data where it lives

The second decision is architectural, and it's the one I'd expect labs to actually adopt for. Claude Science runs on the infrastructure you already have — your laptop, a Linux box, or an HPC login node over SSH — and reaches compute through your own cluster or your Modal account. Large or sensitive datasets stay on the systems they already live on; only the context needed for each step is sent to Claude.

The data boundary: compute and raw data stay put; only minimal per-step context crosses to the model.

This is the correct shape for a tool that wants to touch real research data. Genomic and clinical datasets carry governance constraints that an "upload your files to our cloud" product simply can't satisfy. Running on the lab's own infrastructure sidesteps that, and the session-in-memory design — load a massive dataset once, keep it resident while agents work against it — is a sensible answer to both cost and latency. It also drafts a plan, asks before reaching a new resource, and lets you review or revoke any decision before it submits a job. For anyone who has watched an agent confidently do the wrong expensive thing, that gate is not a nicety.

The actor–critic bet

Running through the whole pipeline is a reviewer agent that inspects outputs, flags incorrect citations, untraceable numbers, and figures that don't match their code, and self-corrects as it goes. This is the actor–critic pattern — one agent produces, a separate one evaluates — and Anthropic is betting it on scientific correctness.

I use this pattern in production and I'm convinced it beats a single agent grading its own homework. But I'd be precise about what it buys you: a critic that shares the generator's blind spots catches formatting and traceability errors well, and subtle domain errors far less reliably. The launch's own examples reflect this honestly. Jérôme Lecoq at the Allen Institute built a roughly 20-skill template that uses actor–critic pairs to draft long-form reviews — turning what could take two years into about ten reviews, many over 100 pages — and his team is now working with domain experts to refine the critic agents. At UCSF, Stephen Francis reports germline analyses for glioma in roughly one-tenth the prior time, but his group independently validated the results. Both are treating the reviewer agent as a multiplier on human judgment, not a replacement for it.

What I'd watch

The honest framing is that this is a beta, shared early on purpose. The questions that decide whether it sticks are empirical: does the audit trail hold up when a skeptical reviewer pulls on it, does the critic catch the errors that actually matter, and do labs trust the system enough to put their proprietary pipelines — saved as reusable skills — inside it. Those answers come from the work, not the announcement. But the underlying instincts — provenance by construction, compute and data that stay put, a built-in skeptic — are the ones I'd have argued for. That's more than most tools in this category can say.