Institute of Science Tokyo's Auctopus Runs a Full Quantum Study With AI Agents

Researchers at the Institute of Science Tokyo report a possible first: an end-to-end scientific study conducted by a laboratory of large language model agents. The system generated research questions, rejected weak proposals, derived a theoretical result, checked it on a supercomputer, and tested its prediction on physical quantum hardware. Human researchers built the apparatus and maintained safety interlocks, while the agents made scientific decisions within the defined research scope and sent approved commands to the instruments.

The team’s paper addresses a persistent failure mode in AI scientist systems: confident conclusions based on insufficient evidence. Its autonomous laboratory lets agents veto proposals, request independent theoretical checks, enforce fixed acceptance criteria, and design experiments that could disprove their own claims. Those controls led the system to reject an initial experiment even though its data appeared to confirm the theory.

Vetoes shape the lab

Auctopus, the open-source Python system behind the study, assigns agents roles modeled on a research group. A principal investigator agent maintains a persistent conversation and supervises the project. Proposer, computational researcher, experimentalist, manager, safety, and reviewer agents work in fresh contexts, exchanging documents through an orchestrator. These information boundaries reduce shared assumptions and automatic agreement between roles.

The proposal gate eliminated 74 of 75 submissions. Recorded reasons included logical errors, previously known results, limited scientific value, and the absence of a feasible experiment. The principal investigator identified one rejected proposal as the strongest starting point, requested revisions, and approved its second version.