$ today_in_10_seconds
Anthropic and Citrini Research pointed an internal Claude model at five scorpion neurotoxins and let it design protein minibinders on its own — 211 of them, sent to the bench for testing. The August demo showed one prompt producing a design. This is the same capability running as a loop, at volume, across a target set.
$ lead_story
Anthropic and Citrini Research use an internal Claude model to autonomously generate protein minibinders against five scorpion neurotoxins, testing 211 designs
An internal Claude model generated minibinder candidates against five scorpion neurotoxin targets without a human proposing each design, and 211 of those candidates went into wet-lab testing. The agentic part isn't the protein chemistry — it's that the model is running the full specify-generate-filter-submit cycle rather than acting as an autocomplete for a structural biologist. That's the same shape as a coding agent that writes, runs, reads the failure, and tries again, except the test harness is a physical assay with a multi-week latency. For builders, the interesting constraint is exactly that latency: agent loops have been cheap because feedback was instant and free, and this is what the pattern looks like when each iteration costs real money and real time.
// why_it_matters: The expensive part of an agent loop is no longer the model — it's the verifier, and bench science is the most expensive verifier there is.
$ bigger_picture
The trend behind the headlines
Most agentic AI progress this year has happened where verification is free: code that compiles or doesn't, tests that pass or don't, a browser task that either books the flight or doesn't. Scientific design is the harder regime, because the model can produce a thousand plausible candidates and the world only tells you which ones worked months later and for a fee. What's actually being tested here isn't whether a model can propose a protein — it's whether its internal ranking is good enough that a batch of 211 is worth a lab's time. If that calibration holds, the bottleneck in a lot of R&D stops being idea generation and becomes assay throughput. That's a very different thing to build a company around than "AI writes your code."
RFdiffusion
The Baker lab's open diffusion model for de novo protein backbone design, and the thing most binder pipelines are still built on top of. Worth reading even if you'll never touch a pipette: it's a clean example of a generative model wrapped in a filtering stack, which is the architecture pattern you want when your agent's outputs are expensive to validate. Pair it with ProteinMPNN for sequence design and you have the open-source skeleton of what the closed systems are doing.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime