Haiqu, a company operating in the quantum software space, released its AgenticOS platform to the public on 7 October 2026. The architecture splits a given scientific problem across a team of specialised agents, each handling a distinct portion of the work, with verification steps built in throughout the process.
People remain a central part of the process. Scrutinising assumptions and conclusions falls to researchers at every stage where a project's continuation is under consideration.
Agentic AI — a framework that has gained considerable traction in recent years — is now being extended into the research domain through this approach. Where task-based workflows drive the architecture, AI agents operate on a fundamentally different footing from chatbot-style systems built around answering one question at a time.
Haiqu describes AgenticOS as a platform where the process opens with a researcher posing a scientific question. Within AgenticOS, the inquiry gets broken down into a web of linked research tasks that the platform renders visually. From that pool of agents, the platform can distribute workloads across a range of functions — scanning prior research, working through mathematical derivations, weighing whether a quantum-based approach is viable, and verifying that outputs hold up under scrutiny.
Proponents describe the goal as building something akin to an on-demand research team, though scientists would retain authority over scientific judgement throughout. Outputs from Haiqu's platform can be put through comparison tests against classical-computing benchmarks, with verification available via other AI agents, formal review procedures, and independent researchers.
One constraint that generative AI has long faced in scientific contexts finds a potential answer here. When a calculation is streamlined through approximation, there is a genuine possibility that the original scientific question and what the model ends up solving drift apart from one another.
Testing on the Hubbard model — a quantum-physics benchmark widely treated as one of the field's more demanding problems — has already been carried out by Haiqu. Mykola Maksymenko, the company's co-founder and chief technology officer, pointed out that converting a research concept into a functional experiment draws on accumulated technical expertise, and that AgenticOS is designed to handle much of that process, freeing researchers to concentrate on the questions that matter most and reach their findings more quickly.
A two-dimensional lattice configuration featuring doping, geometric frustration, and diagonal hopping pathways was the architecture the firm settled on. A 3-by-3 grid formed the basis of a purpose-built experiment, allowing quantum calculations to be checked against an exact classical solution, with the trial then carried out on an IBM quantum computer.
Haiqu's own figures put the drop in quantum circuit depth at 75 percent. According to the firm, the AI-driven system identified a calibration adjustment that aligned the hardware's output with what a theoretically perfect quantum computer running the same algorithm would have returned. The numbers originate with Haiqu and reflect that company's own reported findings, not an externally verified assessment of AgenticOS across the broader quantum research landscape.
Haiqu has indicated that a chemistry paper cleared for presentation at a NeurIPS 2026 workshop dedicated to AI-driven scientific research served as the test case for AgenticOS. Internally testing the brief across ten separate runs, the company's own assessment found that four of those attempts by top-tier AI agents did not adhere to a specific instruction tied to how the experiment was configured.
Under AgenticOS coordination, Haiqu reports that its models followed the agreed-upon protocol and produced a simulated quantum outcome landing roughly 3.4 percent away from an exact-diagonalisation benchmark. Richard Givhan, who co-founded the company and serves as its chief executive, noted that structuring, grounding, and verifying what an AI agent produces carries as much weight in scientific research as advancing the underlying models.
According to Haiqu, a completed research plan gets handed off to the firm's quantum software stack through AgenticOS. Two distinct layers divide its architecture: an SDK layer covering data ingestion, circuit compression, and error mitigation, and a Runtime layer that routes jobs across both simulators and quantum processors.
According to the firm, its platform targets R&D teams at large organisations and university labs, along with researchers from other fields exploring whether quantum methods might address specific technical challenges. Enterprise R&D teams can request access to AgenticOS directly, while academic researchers go through a dedicated application programme the company runs.
