Moving the expertise goalposts: will AI make quantum computing easier?

On 8 October 2026, Pasqal announced an AI-assisted workflow to help researchers take an idea through to a running experiment on its cloud-accessible quantum processing units (QPUs). The Neutral-Atom Agentic Toolkit provides skills for widely used AI coding agents, including Claude Code, Codex, Cursor, and Gemini CLI.

"The bottleneck in quantum computing is shifting," said Pasqal CEO Wasiq Bokhari. "The engineering work that once demanded a researcher's attention can increasingly be automated, allowing researchers to focus more fully on the science." 

Cloud platforms make quantum hardware remotely available, but traditionally a researcher would still have to translate a question or hypothesis into a suitable quantum algorithm or experimental protocol, then into instructions the hardware can execute. That requires a specific combination of software and hardware know-how, informed by the physics of the experiment. The researcher must also assess whether the experiment will survive real-world noise and what the results mean. Pasqal's toolkit aims to automate much of the implementation work.

So how much does the researcher still need to understand when an agent handles that work?

Putting an experiment together

Pasqal describes the intended users of the agentic toolkit as "researchers who can formulate and judge a scientific question and are comfortable with Python and a coding agent". They are scientifically knowledgeable users, rather than complete beginners.

Using the Neutral-Atom Agentic Toolkit, the user can start with a question they want to investigate or they can supply a research paper or patent describing a method they want to implement on quantum hardware. The agent proposes a plan for a neutral-atom experiment, which the user can review and revise. It then writes the code and tests it in computer simulations, including the effects of realistic hardware noise, before recommending a run on real quantum hardware. Afterwards, the agent gathers the results and compares them with the predictions. The user remains responsible for deciding what they mean.

The toolkit builds on the Pasqal team's July 2026 paper, "Lowering the implementation barrier of neutral-atom quantum computing with agentic workflows". In one demonstration, the agent reproduced a published density-wave experiment and had jobs running on hardware in under an hour, after just three messages from the user.

Pasqal’s toolkit is part of a wider effort to use AI to make quantum computing easier to develop and use. Elsewhere in the quantum stack, NVIDIA's Ising Calibration and Google's reinforcement-learning work apply AI close to the hardware, supporting calibration and the control of error-corrected systems. Tools such as IBM's Qiskit Code Assistant models and Classiq's Quantum Agent work nearer the user, helping users write quantum code or automatically generate quantum circuits from higher-level descriptions.

These uses broadly divide between improving hardware for engineers and making code easier to write for programmers. Pasqal's toolkit spans the journey from idea to hardware in a single agent. Its particular appeal is to a scientist who has an experiment in mind but is not a quantum programmer.

Knowing what to trust

For that scientist, much of the practical work of implementation can now be delegated. The agent can handle Pulser (Pasqal's pulse-programming library) and the emulators, work within hardware constraints using calibration data, and manage cloud submission. Pasqal's own assessment suggests that researchers could soon have many more experiments to explore. A separate literature-classification agent assessed 258 of 526 usable papers as implementable on Pasqal’s cloud-accessible QPUs, either directly or with modest adaptation.

Of course, as with other uses of AI, the benefits of this assistance depend partly on whether users can recognise when it has gone wrong. Pasqal’s paper describes how expert review caught two mistakes made by the agent. In one experiment, the agent chose a way of analysing the results that could not establish the magnetic pattern being studied. In a separate graph-colouring experiment, it confidently blamed quantum hardware for invalid results that were actually caused by a software labelling bug.

So agents taking on more of the implementation work may well allow researchers to explore a broader range of experiments without the depth of implementation expertise previously required. Yet easier implementation may make other skills more important: choosing experiments that can genuinely answer a research question, knowing when to challenge an agent’s explanation, and judging whether the results can be trusted. The expertise goalposts may therefore be moving towards scientific judgement rather than specialist quantum-programming skills.

A patent as a recipe for implementation

The third demonstration described in Pasqal’s paper will be of particular interest to IP practitioners. The agent began with Pasqal's own patent application WO2025099250A1, "A method for estimating the chromatic number of a graph". The application describes using neutral atoms to colour a graph so that connected points have different colours. In each round, the system selects atoms representing graph points that are not connected to one another and can therefore share a colour. Those atoms are physically removed, and the process repeats on the remaining atoms for another colour. Repeated runs of this process help estimate the minimum number of colours needed.

The paper notes, however, that “the patent specifies the algorithm but not its realization on present hardware, so the agent must bridge from a protocol description to a working multi-round pipeline”. The agent used the hardware measurement results from each round to identify the points selected for the relevant colour. It then removed those points from the graph in conventional software and sent the smaller remaining graph to the QPU as a new run. The authors conclude that "a patent, read directly, served as a complete and actionable protocol source." They describe patents as a "largely untapped corpus of hardware-ready protocols.”

Those who work in the patent profession understand well the bargain of patent protection: a period of enforceable exclusivity in return for teaching the public how to put the invention into practice. A patent is, in that sense, a recipe. The Pasqal demonstration suggests that AI agents could help readers put those recipes to use, even where implementation needs adapting to the available hardware.

This kind of machine-assisted implementation of patent disclosures raises some interesting questions. In this demonstration, Pasqal were working from their own patent application, but agents implementing protocols disclosed in granted and enforceable third-party patents could face a risk of infringement. Would the agent understand such a risk? Would the user? And if the user doesn’t sufficiently understand the implementation details proposed by the agent, would they actually be able to determine the risk level?

Where expertise matters now

Pasqal’s launch suggests that the expertise needed for useful quantum work is changing. If agents reduce the time and specialist engineering effort needed to move from an idea to a QPU run, implementation may become less of a barrier for researchers who already have access to the hardware.

Progress would then depend increasingly on researchers who can choose worthwhile experiments and properly interpret the output. And if they are going to be encouraged to turn to patents as an input for the agents, perhaps robust IP training will become even more important.

What are your thoughts – will AI agents widen access to quantum computing, or simply move the bottleneck elsewhere?