Haiqu puts AI agent teams to work on quantum research, with human checks in the loop

Haiqu says its AgenticOS system divides scientific research across specialised AI agents, with checks by other agents, formal verification and researchers. In an early quantum-physics test, the company says it cut circuit depth by 75 per cent on an IBM quantum computer, though the result still needs independent validation.

Watch Desk analysis

What happened

The system is designed to take a research question through tasks including literature review, mathematical derivation, assessing whether a quantum approach is feasible and checking results. Researchers remain responsible for examining the assumptions and conclusions.

Haiqu says a test on a 3-by-3 version of the Hubbard model compared the quantum calculation with an exact classical solution. The company reports a 75 per cent reduction in quantum circuit depth and says calibration brought the hardware result into line with an error-free quantum computer using the same method.

In a separate chemistry study, Haiqu says four out of ten runs by leading AI agents failed to follow an explicit instruction about the experimental setup. With AgenticOS overseeing the workflow, the company reports that the models followed the approved protocol and produced a simulated result about 3.4 per cent from an exact-diagonalisation reference. Read the Digital Journal report.

Why it matters

Scientific work can go wrong even when an answer looks polished: an AI system might quietly change an assumption or ignore a constraint, leaving researchers with a result to a different question. Haiqu’s approach aims to make the work more structured and checkable, while using agents to handle parts of the research process.

Shorter quantum circuits could matter on noisy hardware, where errors can accumulate. But the reported results are Haiqu’s own, from specific tests. They do not establish that the system will make quantum research more reliable in general.

Our read

The promising idea here is not an AI scientist with the keys to the lab. It is a team of agents doing bounded work while researchers keep hold of the scientific judgement. The next test is whether independent researchers can reproduce the results on different problems, not whether several agents can agree with one another.

What to watch

• Whether independent teams reproduce the Hubbard-model and chemistry results.
• Whether AgenticOS maintains its checks on larger problems and different quantum hardware.
• How researchers can inspect the system’s assumptions, methods and intermediate work.

Discussion spark: For AI-assisted science, what should count as convincing evidence: reproducible results from independent labs, or a transparent workflow that researchers can inspect?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.