Can teams of AI agents make quantum research more reliable?

Artificial intelligence is becoming sufficiently capable to assist with scientific research, but there is an important difference between producing a plausible answer and producing scientifically defensible work. This distinction is significant when mathematics, simulation and experimental design are involved. Generative AI systems can produce convincing-looking calculations and computer code while still introducing errors. Indeed, recent benchmarking has illustrated that mathematical tasks remain challenging for large language models, highlighting the risks of relying on fluent output as a substitute for verification. Digital Journal has previously examined this problem.

Quantum software company Haiqu is now seeking to tackle this problem through AgenticOS, a new system unveiled on 7 October 2026. Rather than asking one AI agent to carry an entire scientific problem from question to answer, the platform is designed to divide the work between specialised agents and introduce checks along the way. The important component is human involvement. Researchers remain responsible for examining assumptions and conclusions before deciding whether a project should progress.

The concept brings the increasingly popular agentic model of AI into research. AI agents differ from conventional chatbot-style systems in that they can be organised around tasks and workflows rather than simply responding to individual questions. This concept is already spreading into other professional fields. For example, engineering platforms are beginning to use multi-stage agent systems for planning, implementation, testing and review rather than limiting AI to individual coding tasks, as Digital Journal has reported.

With AgenticOS, according to Haiqu, a researcher begins with a scientific question. The system can then allocate agents to activities including reviewing existing literature, conducting mathematical derivations, assessing the feasibility of a quantum approach and checking results. The result is intended to resemble an on-demand research team, although one in which scientists remain responsible for scientific judgement.

AgenticOS displays the investigation as interconnected research tasks. According to Haiqu, results can be tested against classical-computing baselines and checked by other AI agents, formal verification processes and researchers themselves. This addresses an important limitation of generative AI in science. A model could, for example, introduce an approximation that makes a calculation easier without adequately preserving the original scientific question. A result might consequently appear coherent while describing a subtly different problem.

“Turning a promising research idea into a real experiment demands years of specialized, hands-on expertise in highly technical methods, all of which takes time and resources,” explains Mykola Maksymenko, co-founder and CTO of Haiqu. He adds: “AgenticOS takes on much of that labour-intensive work, freeing up researchers so they can spend their time working on the right questions, reaching results much faster without sacrificing accuracy.”

The last point remains something the technology will need to demonstrate across different problems. Scientific reliability cannot ultimately be established by the fact that multiple AI agents agree with one another. Independent benchmarks and human scrutiny remain important.

Putting the system to a physics test

Haiqu says it has already tested the idea on a particularly demanding quantum-physics problem: the Hubbard model. The company selected a two-dimensional doped and frustrated version involving diagonal hopping. It constructed an experiment using a 3-by-3 grid, enabling the quantum calculation to be compared with an exact classical solution, and ran the experiment on an IBM quantum computer.

According to Haiqu, AgenticOS reduced the quantum circuit depth by 75 percent. The company also says the system proposed a calibration that brought the hardware result into line with what an error-free quantum computer using the same method would produce. These figures come from Haiqu and should therefore be viewed as reported experimental results rather than independent validation of AgenticOS across quantum research generally. Nevertheless, circuit depth is an important consideration when working with current quantum computers because real devices remain affected by noise and hardware imperfections.

The difficulty is well established. Quantum information can be disrupted by environmental interactions, crosstalk and imperfect hardware controls, while errors become increasingly problematic as computations grow. Research into the simulation of noisy quantum circuits highlights noise as one of the central obstacles confronting the field. Consequently, shortening circuits without changing the underlying scientific calculation could be valuable because it potentially reduces the opportunities for errors to accumulate.

Putting guardrails around scientific AI

Haiqu further reports testing AgenticOS during a chemistry study accepted at a NeurIPS 2026 workshop on AI for Scientific Discovery. According to the company, when leading AI agents were independently given the research brief, four out of ten runs failed to follow an explicit instruction concerning the experimental setup. With AgenticOS overseeing the workflow, Haiqu says the models adhered to the approved protocol and generated a simulated quantum result approximately 3.4 percent from an exact-diagonalisation reference.

Again, these are company-reported results. They do, however, illustrate the problem AgenticOS is designed to address. An AI system may produce something technically sophisticated while quietly abandoning one of the constraints that made the original research question meaningful.

“Being able to organize, ground and check an AI agent’s work matters as much for scientific research as improving the models themselves,” says Richard Givhan, co-founder and CEO of Haiqu.

From scientific question to quantum hardware

The broader ambition is to link AI-assisted scientific reasoning with the infrastructure required to perform an experiment. Once a research plan has been developed, Haiqu says AgenticOS can pass it to the company’s quantum software stack. Its SDK addresses areas including data loading, circuit compression and error mitigation, while its Runtime handles execution across simulators and quantum processors. The company says its platform is aimed at enterprise R&D groups and university laboratories, as well as researchers in other disciplines investigating whether quantum approaches could solve particular problems. AgenticOS is currently being offered to enterprise R&D teams upon request, while academic researchers can apply to the company’s programme for access. This does not mean the autonomous AI scientist has arrived. The more interesting proposition may be almost the opposite.

Instead of attempting to remove scientists from research, AgenticOS is designed to automate portions of the labour surrounding scientific reasoning while exposing enough of the process for humans to interrogate it. That emphasis mirrors a wider lesson emerging from enterprise AI deployments: human oversight remains particularly important where AI outputs influence consequential decisions, something Digital Journal recently explored.

For quantum research, where a small mathematical or methodological change can alter the problem being solved, traceability could prove as important as speed. The challenge for Haiqu will now be demonstrating that its encouraging initial results can be reproduced independently and sustained as the scientific questions, quantum circuits and research teams become considerably larger.