AI agents do real science, with receipts

A structure in 34 hours

Anthropic's biology research team determined the 3D structure of a vault particle found in prokaryotic cells, specifically within a bacteriophage, the first such structure for a prokaryotic vault. The phage-derived vault is approximately 54 nanometers long and 45 nanometers wide, built from 96 copies of a single phage protein called L05. Vault proteins are present in 29 major groups of bacteria and four of archaea, yet found in less than 1% of species with a sequenced genome, which is why the particle stayed unsolved. m:…5742

Claude was integrated into every stage, according to Anthropic: searching protein databases, designing DNA constructs, and processing microscopy data, with RELION software used for cryo-EM processing. The team's optimized purification method takes approximately 90 minutes, and the first structural model was ready in 34 hours. The team also engineered a larger vault by duplicating segments of the L05 protein, and the structure contains an unidentified small molecule resembling coenzyme A. m:…5742

The findings were submitted to bioRxiv, so this is a preprint, not a peer-reviewed result, and the structure claim should be read with that caveat. But the shape of the work is what matters: the agent did not just suggest an idea, it participated in the pipeline from sequence search to microscopy processing, and the result is a concrete, checkable structure. m:…5742

Sixteen million lines of checked mathematics

Axiom released a Lean formalization of the entire Stacks Project, a reference work for algebraic geometry, spanning 16 million lines of code. The company deployed coordinated swarms of AI agents to translate, prove, and integrate theorems in parallel, and the formalization interface tracks mappings between informal Stacks tags and formal Lean declarations. Formal verification is becoming essential, Axiom argues, as AI systems generate mathematical arguments at a pace beyond human inspection. m:…1851

The process surfaced dozens of statements in the Stacks Project requiring refinement, which is a useful result in itself: machine-checking a reference found places where the reference needed tightening. The Stacks Project was initiated by Johan de Jong in 2005, and de Jong requested that AI-generated comments and pull requests not be submitted to the upstream repository. Axiom's formalization is an independent open-source project it maintains, not an upstream contribution. m:…1851

Axiom had previously released a Lean formalization of the Classification of Finite Simple Groups on October 27, 2026. Two massive formalizations in quick succession suggest the agent-swarm approach scales, at least for the kind of systematic translation work that a reference project requires. The open question is whether the same approach handles novel mathematics, where there is no reference text to translate. m:…1851

Molecules, with experimental validation

MoLRSI, an AI agent for molecular discovery, optimizes molecular design workflows while retaining cross-task memory, and recursively improves its search policy from performance feedback and refinement outcomes. It balances competing biological objectives including antibacterial activity, cytotoxicity, and structural novelty, and uses execution feedback to guide subsequent searches. That loop, memory plus feedback, is what separates it from a one-shot generator. m:…8161

In testing, MoLRSI achieved an experimental hit rate of 15.9% in antibiotic discovery, against 8.7% for expert-led workflows in the same task, and produced eight experimentally validated hits against MART-1 and SARS-CoV-1 peptide-MHC complexes. The paper also reports that Claude Science failed to produce validated hits in the peptide-MHC binder design task. Experimental validation is the key word: these are not predicted scores, they are hits confirmed in the lab. m:…8161

A second paper shows an autonomous agent adaptively designing target-specific multiple sequence alignments for AlphaFold 3, treating the frozen model as a black-box experimental instrument. Evaluated on 1,015 protein-protein interfaces, it raised mean DockQ on FoldBench from 0.465 to 0.612 and on HD439 from 0.482 to 0.562, using on average 60% fewer alignment sequences than the default pipeline, with the largest gains for inter-species heteromers. The agent uses biological hypotheses and self-directed experimentation, refining sequence selection and pairing from predictor label-free signals alone. m:…1946