Lowering the implementation barrier of neutral-atom quantum computing with agentic workflows
Constantin Dalyac, Alexandre Dauphin, Loïc Henriet, Christophe Jurczak
quant-ph, cond-mat.quant-gas, cs.AI
2026-07-28
Pasqal's Claude agent compiled papers into overnight FC1/SA1 QPU runs. Experts caught a wrong observable and a fake hardware leak. 49% of 526 Rydberg papers can run today.
Neutral-atom quantum processors are already bookable in the cloud. Pasqal's FC1 sits in Sherbrooke, Canada; SA1 sits in Saudi Arabia. An external account is enough to submit jobs. Hardware availability is no longer the binding constraint. Between a published protocol and a shot on a QPU sits a long engineering chain: extract the procedure, compile it into device-level pulses, validate it in simulation, navigate the cloud API, and read a physical observable out of noisy bitstrings. Each step has a tool. Stitching them still requires fluency in quantum physics, software engineering, numerical simulation, and hardware calibration. Plenty of theoretically viable proposals stall there.
Pasqal and Quantonation hand that chain to an agentic workflow built on Claude Code (Opus 4.8 and Fable 5), with the researcher in the loop at the decision points. The aim is to shrink the time from paper to QPU, not to let the model invent new physics on its own.
The workflow sits on Pasqal's existing stack: Pulser for pulse design, the matrix-product-state emulator emu-mps, and the Pasqal Cloud SDK. Six skills share one artifact, experimentspec.json. Change the spec, and every downstream stage follows.
Nothing reaches the QPU without passing emulation. A researcher can edit the spec, the register geometry, or the pulse schedule, and the agent regenerates the next round. On these machines the three knobs are atom positions (which set the interactions), the Rabi frequency Ω (transverse field), and the detuning δ (longitudinal field). Each device caps atom number, minimum spacing, reachable Ω and δ, ramp rates, and sequence duration. The agent's job is to squeeze a literature protocol into those boxes.
Three case studies of increasing difficulty, plus a corpus screen.
The first reproduces the 2019 Keesling et al. Kibble-Zurek experiment on a 1D Rydberg chain. The agent concluded on its own that the Z3 phase would force Ω/2π down to 0.087 MHz, so neither a Rabi period nor the nearest-neighbour interaction time fits inside FC1's 6 μs cap; Z4 is further out. The Z2 phase is native: at the 5 μm spacing floor, Rb/a=1.28 sits inside the ordered lobe. The researcher suggested a ring instead of an open chain, and bringing in SA1. FC1 ran a 56-atom ring (2000 shots, correlation length ξ=4.2); SA1 ran a 96-atom ring (1000 shots, ξ=4.2). Both match the original 51-atom chain (ξ=3.9) within statistical error at all 20 lattice sites. Noisy emulation predicted 80% noise retention and an order parameter at 71% of the ideal value. Jobs were on the QPU in under an hour, after three human messages.
The second starts from Guo et al. on frustrated magnetism on a triangular Rydberg array, a theory paper with no hardware recipe. The agent judged the 1/3-filling lobe (δ/Ω=1.22) feasible at full Ω/2π=2 MHz; the 1/2-filling lobe (δ/Ω≃9.5) only marginally so, needing Ω/2π≃1 MHz; the 2/3-filling lobe (δ/Ω=17.8) infeasible, because Ω/2π would have to drop below 0.6 MHz and the adiabatic ramp would outlast coherence. The limiter is FC1's maximum detuning |δ|/2π≤10 MHz. Left alone, the agent kept only the structure factor and the mean magnetization. Mean magnetization cannot tell one pattern from another at the same density. The researcher pulled it back to the staggered magnetization |m| and to rhombic registers with L≡1 mod 3 (N=49 and 100). The 1/3 plateau is clear on the QPU, at roughly half the quantum Monte Carlo thermodynamic-limit amplitude; the missing half is finite-size rounding, diabatic excitations, and hardware noise. This run took 43 human exchanges.
The third source is a Pasqal patent on graph coloring by peeling Rydberg independent sets. The agent built a multi-round measurement-feedback loop with an open batch and MappableRegister, beam search of width 3, 200 shots per round. On 75 unit-disk graphs with N from 15 to 80, the classical chromatic number χ (DSATUR) stays at 4-5. At N=20, all 10 instances hit the exact χ. At N=80 the gap is 1.44±0.50, never above 2, overall mean 0.84.
A second, lighter agent screened 633 Rydberg-array theory papers. Of the 526 it could classify:
| Class | Papers | Share |
| Runnable on today's QPU | 225 | 42.8% |
| Runnable with modest adaptation | 33 | 6.3% |
| Needs XY interactions | 141 | 26.8% |
| Needs local addressing | 87 | 16.5% |
| N>100 or other | 40 | 7.6% |
That is 49% on present hardware. XY plus local addressing account for about 85% of the rest. The classifier used five coarse rules (N≤100, global drive only, Ising/blockade, no mid-circuit measurement, computational-basis readout), not a detailed simulation.
For anyone who wants to run analog simulation or combinatorial optimization on a neutral-atom machine, this compresses writing Pulser, checking calibration, queuing, and harvesting from days into an overnight loop. The skill shape also shows the quantum cloud stack is now programmable enough for an agent to plug in.
A successful run is not a correct measurement. Example A needed three messages because the target was a known result. Example B measured the wrong observable. Example C blamed Rydberg-blockade leakage for what was a coordinate-to-bitstring ordering bug. Both failure modes produced plausible plots. Only a domain expert stopped them. The bottleneck has moved from "can we run this" to "are we measuring the right quantity, and is the story physics."
The corpus scan also gives hardware a demand-side number. Unlocking the largest remaining slice of the literature means XY interactions and local addressing, not pushing atom count past 100.
The paper is unusually honest about the failures. Several caveats still apply.
The skills are welded to Pasqal's Pulser and Cloud SDK. A different neutral-atom machine, or a superconducting platform, needs the interface layer rewritten. One of the three demos comes from an in-house patent, and the coloring numbers are on sparse unit-disk graphs whose χ sits at 4-5. That does not transfer to hard coloring instances.
The literature screen is a back-of-the-envelope filter. 107 PDFs were unreadable and dropped. The five-rule tree labels a paper by the first rule it fails. No human audit rate is reported. Read 49% as "looks runnable after a coarse screen," not "258 experiments have been reproduced."
This is also a vendor paper about the vendor's own cloud and the vendor's own agent. The hardware data did come through standard external cloud accounts. The comparison, though, is mostly how long a human would take to do the same plumbing, and those hours are not in the paper. There is no ablation against a no-agent script pipeline. What is solid: the human stays in the loop, and has to.