A scientific calculation can run perfectly and still answer the wrong question. If an AI agent quietly changes a molecule’s geometry or drops electrons from a simulation, working code and a convincing graph can hide a broken experiment.
That is the problem Haiqu is targeting with its October 7 announcement of AgenticOS, a system that coordinates teams of specialized AI agents across quantum research projects. The platform is designed to carry an idea through literature review, mathematical analysis, experiment design and execution while keeping scientific assumptions available for researchers to inspect and approve.
Available now to enterprise R&D teams upon request, AgenticOS builds on Haiqu’s platform announced in May. Its central proposition is that useful scientific automation needs a way to preserve the experiment’s meaning across every stage of the work.
Turning a research question into a reviewable workflow
According to Haiqu’s AgenticOS product page, users can begin with a question, papers, data or an existing idea. The system helps establish objectives, constraints and success criteria, then organizes the investigation into a research graph. Nodes represent teams of scientific agents, with dependencies, decisions and research artifacts connecting the work.
Researchers can examine outputs, comment on assumptions, redirect a module and approve important decisions before downstream tasks proceed. The system draws on a knowledge base of quantum theory, algorithms and industry applications, together with the project’s accepted decisions and artifacts.
That structure matters because a research project can fail long before its final calculation. Choosing an approximation changes what is being modeled; losing an earlier constraint can make later results incomparable. A visible workflow gives scientists places to catch those changes, instead of trying to reconstruct them from a finished answer.
In the announcement, Haiqu says the agents handle literature review, first-principles derivations, quantum feasibility analysis and result checking. Validation combines classical baselines, tests, critic agents and human review, with formal verification. Scientists determine which assumptions are locked, where autonomous work is appropriate and which steps require sign-off.
Co-founder and CTO Mykola Maksymenko said the system is intended to reduce the labor required to turn a promising research idea into an experiment, allowing scientists to devote more attention to selecting and investigating the right questions.
A chemistry test exposes the difference between code and science
Haiqu’s chemistry case study examines proton transfer in the Zundel cation, H₅O₂⁺: two water molecules sharing an extra proton. The task was to calculate the energy barrier as that proton moves between them, under a fixed scientific setup.
Ten standalone AI runs received the same brief. All produced working code, but four excluded required electrons, and others altered the geometry or proton’s path. AgenticOS preserved the agreed protocol, including all 20 electrons and a fixed oxygen–oxygen separation.
Its simulated sample-based quantum diagonalization, or SQD, calculation produced a barrier of 554.8 millielectronvolts, versus a 574.3-millielectronvolt full configuration interaction reference—about a 3.4% difference. SQD combines sampling of electronic configurations with a classical energy calculation.
Those numbers describe one selected molecular model. The reference is exact within that model’s computational assumptions; the result does not establish universal accuracy or quantum advantage. The more immediately relevant finding is that preserving the protocol made the calculation meaningful to compare.
Haiqu says the work was accepted at a NeurIPS 2026 workshop focused on AI for scientific discovery. It remains a company-reported demonstration, rather than a broad independent benchmark of research reliability.
From a controlled simulation to quantum hardware
The announcement also describes a separate experiment involving the Hubbard model, a framework for studying interacting electrons. Haiqu investigated a doped, frustrated two-dimensional version with diagonal hopping, relevant to research on copper-based high-temperature superconductors.
AgenticOS developed an experiment on a three-by-three grid and ran it on an IBM quantum computer. The small grid allowed comparison with an exact classical solution. Haiqu reports a roughly 75% reduction in circuit depth and a proposed calibration that aligned the hardware output with an error-free execution of the same method.
Circuit depth measures the layers of operations needed to execute a quantum circuit. Reducing it can help on today’s noisy processors, where errors accumulate during computation. Agreement with an ideal execution is useful evidence about the method and hardware workflow, but it does not mean the experiment has solved superconductivity or surpassed classical computing at that scale.
The software stack behind the agents
AgenticOS sits above Haiqu’s quantum software infrastructure. The announcement describes the SDK as providing data loading, circuit compression and error mitigation, while Haiqu Runtime manages execution on simulators and quantum processors.
Haiqu’s academic platform page adds practical detail: researchers can bring Qiskit code into the SDK, track jobs and preserve backends, calibration information, random seeds and compilation settings. Those records help teams reproduce an experiment and understand why two executions differ.
The platform also estimates quantum processing costs before execution and uses batching and caching to reduce unnecessary runs. Such capabilities connect the agent workflow to the operational constraints of a real research budget.
Enterprise teams can request AgenticOS access. Academic researchers can apply for free platform access through Haiqu’s Academic Program and seek AWS credits for classical and quantum workloads; the company notes that credits require approval and are not guaranteed.
For research teams, the promising part of the announcement is the continuity between question, assumptions, code and evidence. Faster calculations are valuable only when scientists can determine what was actually calculated. AgenticOS makes that requirement a central part of its approach to AI-assisted quantum research.
Credit: Source link
