A biotech staffed entirely by agents
-
Nº CII
- Date
- 18 Sep 2026
- Issue
- 102
- Stories
- Eight
- Editor
- ARC
Agents got an org chart today, and frontier biology capability got a bouncer.
Science paper runs drug discovery with thousands of agents
Tens of thousands of agents run a simulated biotech end to end in VirtualBiotech, published in Science and announced by Stanford's James Zou. Agent populations coordinate across the full pipeline, from target discovery through clinical development, instead of pointing one model at one step. Scale is the claim — most agent demos in drug discovery have been narrow and single-task, and they rarely survive the handoffs between stages. Landing in Science moves multi-agent drug discovery from engineering demo to peer-reviewed result, and gives the field a reference architecture it now has to argue with.
Anthropic opens a verification path for life-science users
Anthropic launched a Life Sciences Verification Program, a formal route for life-sciences organizations to verify themselves with the company. Frontier AI vendors have spent two years restricting biology-adjacent capabilities to limit misuse, and legitimate research kept getting caught in the same net. A credentialing path turns blanket caution into a check on who you are, which makes verified status a live distinction in what biology work frontier models will do at all. What verification unlocks in practice is what the program's details have to answer.
Liquid AI puts longevity models on Cell's cover
Liquid AI claims top scores on biological longevity tasks in a Cell cover article, using instances of its Liquid Foundation Models, a model family built on an architecture other than the transformer that underlies most LLMs. A non-transformer stack on the cover of Cell is a real datapoint in an argument that has mostly run on benchmarks: whether architecture choices made for general modeling transfer to aging biology, where data is sparse and endpoints are slow. Company-reported wins still want outside replication.
Anthropic measures Claude's lift in biomolecular modeling
Claude helps with biomolecular modeling , Anthropic says in a research writeup on where its general-purpose model provides uplift. Vendor-run evidence, but it sharpens a live question: how much structural and molecular work a general model carries before a purpose-built tool is required.
LLMsFold checks generated drug candidates against physics
LLMsFold runs generated molecules through biophysical simulation for de novo drug design, testing candidates against physics rather than model confidence alone. Coupling generation to simulation tightens the standard for what counts as a designed molecule instead of a plausible string.
TorchCraft inverts a structure predictor to design binders
TorchCraft turns a structure predictor backwards into a binder generator, doing design with the all-atom prediction model itself instead of training a separate generative one. One model for both jobs collapses the two-tool workflow most binder pipelines still run end to end.
OpenAntigens indexes the surfaceome for antigen design
OpenAntigens maps antigen constructs across the human cell-surface and secreted proteome in a structure-aware database. Construct design has run largely on group-level know-how and one-off modeling; a shared reference makes those choices checkable between labs.
CliniCIRCA builds mental-health timelines from raw EHR notes
CliniCIRCA rebuilds patient journeys from unstructured EHR narratives, assembling longitudinal mental-health timelines through a modular chain of language-model steps. Ordering psychiatric notes into usable clinical histories is the bottleneck that has kept this record type out of scaled research.
Reply with your discoveries. A human reads them. Forward freely.
|