Scientific agents need a stage-conversion ledger
In short: Anthropic's protein-binder campaign shows why scientific agents should be measured by conversion, time and cost between evidence gates.
On 20 October 2026, Anthropic reported that an internal Claude research model generated 318,484 protein designs for five scorpion toxins in 64 hours. The campaign used roughly 10,000 GPU hours. Two laboratories tested 211 selected designs, and 17 met Anthropic's binder criteria.
The headline number is 318,484. The operating number is the conversion between each evidence gate. My rule is simple: a scientific-agent programme needs a stage-conversion ledger that records entrants, exits, elapsed time and cost at each gate. Candidate volume and autonomous hours measure activity, not scientific outcomes.
Anthropic's 318,484 designs became 17 binders
1/ Anthropic's research account says Claude used 12 design tools, co-folded 553,052 design-toxin pairs and worked for 24 hours before the first human follow-up. Four follow-ups shaped the 64-hour campaign. “Autonomous” here still includes a defined prompt, supplied tools, human direction and laboratory partners.
2/ Claude selected 180 designs for testing. Anthropic's team added another 31 with Claude's help, producing the 211 candidates sent to Adaptyv Bio and Twist Bioscience. The public data release includes both laboratories' results, affinity measurements and off-target checks.
| Evidence gate | Count reported |
|---|---|
| Designs generated | 318,484 |
| Designs tested in laboratories | 211 |
| Designs meeting binder criteria | 17 |
| Multitoxin designs binding more than one target | 0 |
| Designs tested for cellular neutralisation | 0 |
| Designs tested for animal efficacy | 0 |
3/ The 17 binders are a laboratory result within the stated criteria. They are not an antivenom. Anthropic has not yet tested whether the proteins neutralise venom in cells or protect animals, and its strongest candidates were tested against recombinant toxins rather than venom purified from scorpions.
The zeroes identify the next scientific bottleneck
4/ None of the designs intended for multiple toxins met the binder criteria for more than one toxin. That zero should remain visible. Folding it into an aggregate “17 successes” would hide the exact capability the universal-cocktail idea needed.
5/ A later gate can reverse the apparent value of an earlier one. A 2025 Nature study on snake-toxin binders moved beyond binding into cellular neutralisation and mouse experiments. It showed that those gates are possible in related work; it does not supply missing evidence for Anthropic's scorpion designs.
6/ The stage-conversion ledger should therefore carry separate denominators: generated, computationally screened, selected, expressed, bound, functionally neutralised, effective in animals, safe, manufacturable and accessible. Each transition gets its own start time, completion time, compute, laboratory cost, human intervention and failure reason.
My stage-conversion ledger for scientific agents
7/ This is labelled editorial judgment, not a drug-discovery programme I ran at Trigent or for a client. I would use the same ledger for an enterprise agent: proposals generated, proposals passing deterministic checks, cases sampled by reviewers, changes accepted, changes surviving production and outcomes retained after 30 days.
8/ The scorecard stops teams from celebrating the easiest denominator. If generation rises tenfold while accepted production outcomes stay flat, the agent did not deliver ten times more value. It moved the bottleneck. Fund and staff the next gate before paying for more candidate volume.
Anthropic's figures and resource use are first-party. I could inspect the released 211-design dataset and the published methods, but I could not independently reproduce the private model session, laboratory measurements or reported timeline. The ledger makes that boundary part of the result instead of burying it in a footnote.
What's in it for you
- Name the evidence gates before the next autonomous run begins.
- Report conversion, elapsed time and cost at every transition, including zeroes.
- Increase generation capacity only when the next gate has review and execution capacity.
Count what survives each evidence gate; raw AI output is only the queue entering the first one.
