AI Agent UpdateClaude designed 17 scorpion-toxin antidotes in 64 hours — mostly by itself
One prompt, four follow-ups, 318,000 designs, two independent labs. The AI-agent era's cleanest science demo yet — and the data is all public.
The News
Anthropic published a research post on Oct 20 describing a strikingly hands-off experiment. Working inside Claude Science, its research workbench, Claude was given one initial prompt plus four follow-up messages — and then ran a 64-hour campaign largely on its own, designing protein 'binders' against scorpion neurotoxins. Binders are small engineered proteins designed to latch onto a specific target; latch onto a toxin tightly enough, and you have the starting point for a treatment.
The scale is the headline: Claude generated 318,484 candidate designs and screened 553,052 design–toxin pairings using 12 different tools, narrowing the field to 211 candidates worth making in a real lab. Those 211 were tested at two independent labs — Adaptyv Bio and Twist Bioscience — and 17 of them actually bound their targets, covering 4 of the 5 scorpion neurotoxins aimed at. Two bound the Mexican Cn2 toxin at single-digit nanomolar affinity, which in plain terms means they grip it extremely tightly.
Anthropic is careful about what this is not. Binding a toxin in a test is not the same as neutralizing it in a living body: the company stresses it has NOT tested neutralization in cells or animals, so this is not an antivenom claim. What it did do is release everything — all 211 designs and the binding data are public, so any lab in the world can check the work or build on it.
One prompt, four follow-ups, 64 hours: Claude came back with 17 candidate antidotes — and published all of its homework.
Why It Matters
Most AI-agent demos are software doing software: writing code, browsing, booking. This one crosses into the physical world with a complete loop — hypothesis, design, screening, and independent wet-lab verification — run from five human messages over 64 hours. Protein design work that would occupy a specialist team for months was compressed into a weekend of compute, and the results survived contact with two outside labs.
For Anthropic, Claude Science is the product signal: agents that can operate scientific tooling end-to-end are a different business from chatbots, aimed at pharma and biotech budgets. And publishing all 211 designs is a credibility play — in a field drowning in demo videos, releasing checkable data is how a lab says 'verify us.'
Analyst's Take
The ticker note, plainly: Anthropic is a private company — there is no Anthropic stock. The tickers on this story, Google ($GOOGL) and Amazon ($AMZN), are read-throughs: they are Anthropic's biggest cloud investors, so Claude's expanding usefulness lands partly on their platforms. The investment story here is the agent platform, not a scorpion drug.
What makes this the cleanest agent demo yet is the restraint around it. Five prompts. Independent labs. A public dataset. And a headline caveat — no cells, no animals, no antivenom claim — stated by the lab itself rather than discovered by critics. The thing to watch is the follow-through: whether other labs replicate the binders from the public data, and whether the next campaign closes the loop into cell testing. If agents can reliably run design-to-data science at this tempo, the bottleneck in early drug discovery stops being ideas and becomes lab benches — a very different, very fundable problem.
By The Numbers
FAQ
What is a protein binder?
A small protein engineered to attach tightly to one specific target — here, scorpion neurotoxins. A binder that grips a toxin strongly is the first step toward a treatment that could block the toxin's effect, though many more steps (cell and animal testing) come after.
Is this a working antivenom?
No — and Anthropic says so itself. The 17 binders were shown to attach to their toxins in lab tests, but neutralization has not been tested in cells or animals. It is a validated starting point, not a medicine.
Who checked Claude's work?
Two independent labs — Adaptyv Bio and Twist Bioscience — physically tested 211 of Claude's designs. Anthropic also released all 211 designs and the binding data publicly so anyone can verify or extend the results.
How much human input did it take?
Remarkably little: one initial prompt and four follow-up messages, inside Anthropic's Claude Science workbench. Claude ran the 64-hour campaign — 318,484 designs, 553,052 screenings — using 12 tools largely autonomously.
Source: Anthropic via The Decoder