Key news
Quantum software company Haiqu unveiled AgenticOS on October 9, 2026: an AI agent system built for scientific discovery, aimed at letting AI genuinely help solve the hardest problems in physics while ensuring that what it computes can be trusted.
The problem it solves
- When general-purpose AI agents do scientific computation, they quietly "tamper with the science": for example, moving molecular geometries or sneaking in unapproved approximations, so that results look right but answer the wrong question
- AgenticOS uses "multi-team cross-validation": starting from a researcher's question, teams of agents handle literature review, first-principles mathematical derivation and quantum feasibility assessment, and the work is then checked layer by layer through classical baseline tests, critic agents, human review and a formal verification layer
- Every assumption and conclusion can be inspected: what is locked, what runs automatically and where a human must sign off are all clearly laid out
Test results
- In a chemistry study accepted to the NeurIPS 2026 AI for Scientific Discovery workshop, without AgenticOS, 4 of 10 agent runs ignored explicit experimental setup instructions and produced very different answers; with AgenticOS, the agents followed the prescribed procedure throughout, and the simulated quantum results deviated from the exact diagonalization reference value by about 3.4%
- It has already been used on a hard physics problem: the two-dimensional doped frustrated Hubbard model, aimed at studying copper-based high-temperature superconductivity. Starting from a scientist's idea, it built an experiment on a 3×3 lattice and ran it on an IBM quantum computer, reducing circuit depth by 75%; after calibration, the hardware results matched what an error-free quantum computer would be expected to produce
Source: read the original. This text is a summary for reference only.
