Claude Science Is Workflow, Not Magic

The Short Version

Claude is moving into the lab.

On June 30, Anthropic introduced Claude Science, a beta workbench for scientific research. It runs on macOS, Linux, remote machines, and HPC login nodes. It connects to tools researchers already use, including PubMed, Jupyter, R, structural biology viewers, genomics databases, chemistry tools, and compute infrastructure.

That sounds like a science chatbot.

It is not just a science chatbot.

The interesting part is that Anthropic is not mainly selling a new miracle model here. Claude Science is an environment. It has more than 60 curated skills and connectors, can spin up specialist agents, produces reproducible artifacts with code and message history, and includes a reviewer agent that checks citations and calculations.

That is the shift.

AI labs are no longer only competing on who has the smartest model in a blank text box.

They are competing to own the workflow.

For software, that looks like Codex, Claude Code, Cursor, and all the agentic coding tools that sit inside repos, terminals, issues, and pull requests.

For science, it may look like Claude Science, OpenAI's Prism, GPT-Rosalind, DeepMind's biology stack, and whatever comes next.

The model still matters.

But the product is becoming the model plus tools, data, compute, permissions, audit trails, and a domain-specific interface that knows what real work looks like.

Very glamorous. We reinvented the lab notebook and gave it subagents.

Update: Claude found something. Biology still gets the last word

Anthropic now has a much stronger example of the workflow in action.

On September 23, the company said Claude agents had identified a previously uncharacterized enzyme system in bacteriophages. Anthropic calls it array-associated reverse transcriptases, or ART. The system combines a reverse transcriptase, a partner gene, and a long array of repeating DNA sequences that looks somewhat like a CRISPR array.

That is a real scientific lead.

It is not a new CRISPR tool.

Anthropic does not yet know ART's biological function. There is no known DNA-cutting enzyme attached to it. The preprint has not been peer reviewed, and the result has not been independently replicated. As Nature reported, the evidence that ART behaves like CRISPR is still scant.

So the useful headline is not "Claude invented gene editing."

It is:

Claude helped find a weird thing worth testing.

That is less cinematic.

It is also much closer to how discovery begins.

The 950-agent number needs a workflow around it

The computational search was substantial.

According to Anthropic's preprint, 949 Claude sessions completed 119 tasks across 21.5 hours of wall-clock time, using 215.6 million tokens. The agents surveyed reverse-transcriptase loci across 1.9 billion protein clusters, recovered roughly 200,000 RT clusters, investigated thousands of nearby protein families, and produced 19 reports for human review.

One agent noticed an unusual repeating pattern in raw DNA beside a reverse-transcriptase gene. The original research brief had asked the system to look for partner genes, not non-coding repeat arrays. The agent followed the anomaly, opened more work, compared it with known systems, and promoted the result for review.

That is the impressive part.

The agents did not simply retrieve a fact that Anthropic had hidden in a document. They ran a large computational campaign and noticed a pattern outside the most obvious path.

But the autonomy had boundaries.

Humans chose reverse transcriptases as the research target. Humans wrote the brief. Humans built the harness and selected the model. Humans reviewed the reports. Human scientists performed all physical lab work. Humans interpreted the evidence and decided what could responsibly be claimed.

The computational campaign ran without human intervention.

The scientific process did not.

Both statements matter.

Hypotheses are getting cheaper than experiments

Claude gathered roughly 200,000 RT clusters and narrowed the campaign to a small set of candidates.

That funnel is the product.

AI can now produce more plausible hypotheses than a lab can physically test. It can search a database all night, open hundreds of parallel investigations, write candidate reports, rank them, and preserve the transcript. The scarce resource moves downstream: expert attention, assay design, wet-lab capacity, controls, replication, and the willingness to kill an exciting idea when the evidence does not cooperate.

This changes what a good scientific agent should optimize for.

Not the number of hypotheses.

Not how confident the reports sound.

Not how many agents were involved.

The useful metric is how efficiently the system moves from a large search space to a small set of testable, well-supported candidates without hiding the rejected paths.

The graveyard matters.

Anthropic's agents investigated 17 candidate partner families. Most were rejected or set aside as annotation artifacts, known systems, or ordinary genomic neighbors. That is healthy. A scientific workflow should make the failures inspectable instead of polishing only the winner into a launch post.

"AI discovered" needs a methods section

This result gives us a better way to read future AI-science claims.

Ask which step the AI actually performed.

Did it search literature?

Did it generate a hypothesis?

Did it notice a pattern in primary data?

Did it design the experiment?

Did it operate the equipment?

Did it interpret the result?

Did an independent lab reproduce it?

"AI discovered X" can describe any combination of those steps. Collapsing them into one verb makes a workflow look like magic and makes scientific accountability disappear at exactly the wrong moment.

For this result, the strongest fair claim is specific: Claude agents autonomously surfaced an unusual genomic pattern and drove the computational analysis that initiated the discovery. Humans did the bench science. The function remains unknown.

That is still important.

It just is not the same as an autonomous scientist completing the full loop.

The lab is now part of the AI product

Anthropic did not stop at releasing Claude Science software.

It built a molecular-biology lab.

That is the strategic signal.

If frontier models can generate biological hypotheses at scale, the company that owns only the model still depends on somebody else's instruments, samples, protocols, queue, and judgment to learn which ideas survive contact with reality. A wet lab closes that loop faster.

It also creates a new conflict of interest.

The company built the model, coordinated the agents, employed the scientists, ran the experiments, wrote the preprint, and published the launch post. None of that invalidates the result. It does mean outside review and independent replication matter more, not less.

The next trustworthy layer is not another reviewer agent from the same system.

It is another lab.

What builders should copy

Most teams are not about to build a biology facility.

They can still copy the useful architecture:

• keep the full agent transcripts and code
• show the funnel from all candidates to the selected result
• preserve rejected hypotheses and the reason each was rejected
• separate model observations from human interpretations
• make every computational result reproducible from primary data
• define which steps require expert approval
• publish negative results, not only launch-worthy successes
• invite independent reviewers before turning a lead into a product claim

This is the same lesson Claude Science suggested in June, now with much better evidence.

The agent can make the search broader and faster.

The workflow has to make the result narrower and more trustworthy.

Why Claude Science is the right story

There was a louder version of this story available.

The Verge reported that Anthropic also wants to develop drugs of its own, starting with neglected diseases. That is splashier. It is also easier to overstate.

Drug discovery is not a neat software problem. A model can help search literature, suggest hypotheses, analyze data, design molecules, and road-test ideas. Then biology does what biology does: refuses to be impressed by your launch post.

Candidates still need experiments. Experiments need labs. Labs need money, people, controls, failed attempts, animal studies, clinical trials, manufacturing, regulators, and time. The Verge spoke with experts who made the boring point that no AI-designed drug has yet made it through clinical trials and FDA approval to reach patients.

So yes, Anthropic's drug ambition is worth watching.

But the more useful near-term story is the workbench.

Claude Science is interesting because it points to where AI products go after chat. Not toward one omniscient model that answers every scientific question from nowhere, but toward agents embedded inside the messy environment where scientific work actually happens.

Science is not just "know biology."

It is read the paper, check the method, find the dataset, write the script, run the pipeline, inspect the figure, notice the weird batch effect, rerun the analysis, cite the source, explain what changed, and make the work reproducible enough that someone else can tear it apart later.

That is workflow.

And workflow is where agents become real.

The product is the environment

Anthropic's announcement is very explicit about the problem: researchers live across fragmented tools.

PubMed is over here. Jupyter is over there. R is somewhere else. The cluster has its own rituals. Chemical structures need one viewer. Protein structures need another. Genomic tracks need another. The source data has its own schema, the paper has its own references, and the final figure needs to be defensible six months later when everyone has forgotten how it was made.

This is exactly the kind of work a normal chatbot is bad at.

Not because the model cannot write useful text.

Because the model is outside the system.

It does not naturally know which file produced the figure, which environment ran the code, which database query pulled the gene list, which citation supports the claim, or whether the number in paragraph four still matches the analysis after the latest rerun.

Claude Science tries to pull the agent into that system.

Anthropic says outputs include the exact code and environment used to produce figures, a plain-language explanation, and the full message history. The app can run on a lab's own infrastructure so sensitive datasets do not need to leave the systems they already live on. It can ask before reaching new resources. It can submit jobs to compute systems. It can fork a session to compare approaches.

That is not a better answer box.

That is an operating layer.

The reviewer agent is the tell

The most important feature may be the least magical one: the reviewer agent.

Claude Science includes a separate agent that checks citations, calculations, and whether figures match their underlying code. Anthropic says it can flag and correct errors as work progresses.

Good.

Also: be careful.

TechCrunch made the obvious but important point that the checker is still part of the same AI system, not an independent source of truth. A reviewer agent can catch a lot of sloppiness. It can also miss shared blind spots, inherit the same bad assumption, or produce the comforting feeling that something has been checked when it has only been checked by another model-shaped process.

That does not make it useless.

It makes it a control, not a guarantee.

This is the difference between AI theater and useful AI infrastructure. The point is not "Claude reviewed it, therefore it is true." The point is "the system made the chain of work more inspectable, so humans can review it with less archaeology."

That matters.

Science does not need more confident prose.

It needs provenance.

Where did the claim come from? Which code produced the result? Which dataset was used? Which version of the pipeline? Which citations are real? Which numbers changed after the final analysis? Which assumptions are still hand-wavy?

If AI helps answer those questions, it is useful even when it is not discovering a drug.

The science race has different strategies

Anthropic is not alone here.

OpenAI introduced Prism in January as a free, AI-native workspace for scientists to write and collaborate on research, powered by GPT-5.2. That product starts closer to the paper: drafting, revision, citations, equations, LaTeX, collaboration, and manuscript context.

Then in April, OpenAI introduced GPT-Rosalind, a purpose-built life sciences model for biology, drug discovery, and translational medicine, available through a trusted access program for qualified customers.

Anthropic's bet is shaped differently.

Claude Science is available in beta to Claude Pro, Max, Team, and Enterprise users. It is less "here is a special biology model" and more "here is a scientific workbench around Claude, tools, compute, and agents."

Google DeepMind has yet another advantage: it owns foundational science systems like AlphaFold and related biology models that other companies can only call into as tools.

So the competition is not one-dimensional.

One lab offers a writing workspace.

Another offers a specialized model.

Another offers owned scientific foundation models.

Anthropic offers a workflow environment around a general model with domain tools and audit trails.

This is where AI gets more interesting than benchmarks.

Different domains will not all want the same product shape. A lawyer, scientist, software engineer, investment analyst, designer, and doctor may all use "AI agents," but the actual agent should not feel the same. The interface, constraints, audit trail, data access, approvals, and failure modes are different.

That is the lesson for builders.

Do not just wrap a chat box around a model and call it vertical AI.

The vertical is the workflow.

The risk is scientific laundering

There is a dark version of this.

If the workbench is good, it can make weak work look more professional.

It can generate clean figures, fluent manuscripts, plausible citations, neat explanations, and long evidence tables. That is helpful when the underlying work is solid. It is dangerous when the underlying work is thin.

Science already has incentives that reward polish: publishable writing, tidy visuals, strong narratives, and confident framing. AI can amplify that. A paper can look more complete before it is more true.

That is why the boring pieces matter so much.

Source code. Environment history. Message history. Data provenance. Citation checks. Human review. Reproducible artifacts.

Those are not admin details.

They are the whole trust model.

If Claude Science becomes a prettier way to produce unreproducible claims, it is bad.

If it becomes a more inspectable way to move through tedious scientific work, it is valuable.

The line between those two futures is not model intelligence. It is product design and lab culture.

What this means for AI builders

The practical takeaway is simple:

Agents need homes.

The blank chat interface is useful, but it is not where most serious work lives. Serious work lives in existing environments: repos, notebooks, docs, channels, dashboards, tickets, databases, clusters, CRMs, design files, patient records, spreadsheets, and internal tools.

An agent becomes more valuable when it can inhabit one of those environments with the right permissions and the right memory.

It also becomes more dangerous.

So the product questions get less glamorous and more concrete:

• What tools can the agent use?
• What data can it see?
• Which actions need approval?
• What does it log?
• Can a human reproduce the result?
• Can the user revoke access?
• Does the agent know when to ask for help?
• What happens when the reviewer agent and the worker agent are both wrong?

Those are not secondary details.

They are the product.

The AI industry spent years asking, "How smart is the model?"

Claude Science is another sign that the next question is, "Where does the model work?"

What this means for scientists

For scientists, the right posture is neither panic nor worship.

Do not treat Claude Science as an autonomous researcher.

Do not dismiss it as a fancy autocomplete either.

Treat it like a junior collaborator with weird strengths: tireless literature search, decent scripting, fast first-pass synthesis, useful figure iteration, good checklist memory, and occasional terrifying confidence.

Give it work where auditability matters.

Make it show its code. Make it cite sources. Make it preserve the chain from data to figure. Make it explain what changed between runs. Make it separate "I found this" from "I infer this." Make it easy for a skeptical human to say no.

That is the real promise.

Not replacing the scientist.

Reducing the amount of scientific work that disappears into tool-switching fog.

The bottom line

Claude Science will not make drug discovery easy.

It will not remove the need for experiments. It will not turn every lab into an autonomous discovery machine. It will not make biology stop being rude to beautiful theories.

But it shows where AI products are going.

The future is not just bigger models answering from a distance.

It is agents moving into the working environment, using the tools, preserving the audit trail, and helping people move through long, fragile workflows without losing the thread.

For software, that meant coding agents.

For science, it may mean the lab workbench.

The important question is not whether the AI can sound like a scientist.

The important question is whether it can make scientific work more inspectable, reproducible, and less painfully fragmented.

That is less magical than "AI discovers drugs."

It is also much more believable.