AI agent system scores up to 100/100 in quantum simulation tests

An AI agent system for running quantum simulations scored between 97 and 100 out of 100 in expert assessments across five repeated tasks, according to work published in Reports on Progress in Physics. The system, called El Agente Cuántico, handled problems from molecular-energy calculations to quantum error correction, and its results appear in a 2026 paper by I. Gustin and colleagues (Rep. Prog. Phys. 89 077602).

Quantum simulations are computer models used to study systems governed by quantum mechanics (the physics of matter and energy at atomic scales). Researchers rely on them to calculate molecular energies, investigate candidate materials, model energy transfer and decoherence (the loss of quantum coherence through interaction with the environment), test quantum-computing algorithms, and predict how quantum systems evolve over time. The outputs often cannot be obtained analytically or experimentally. The cost is a heavy software burden: these calculations typically demand specialist numerical methods and fluency in multiple software tools.

How does the agent system work?

El Agente Cuántico applies an agent-based approach to this workflow. A coordinating agent delegates work to specialist agents, each associated with a different software package. The agents plan multistep tasks, consult documentation, run programs, examine the results they produce, and revise their approach when needed. The output is code, numerical results and figures — the working artifacts a researcher would otherwise assemble by hand.

The agent pattern is not specific to physics. Experimental systems such as SciExplorer have inferred equations and Hamiltonians (the operator that encodes a system's total energy) from simulated observations. Related agent-based approaches are also being explored in mathematics, including the automated search for proofs.

What was measured?

The researchers demonstrated the system on six classes of problems:

• molecular-energy calculations
• entangled states
• dissipative dynamics
• time crystals
• quantum error correction
• hardware-resource estimation

In repeated tests of five tasks, average expert-assessment scores ranged from 97 to 100 out of 100. That figure is the measured result. The authors also disclose the failures: a small number of runs departed from the instructions, for example by calculating the wrong observable (the quantity being measured) or preparing an initial state with the wrong sampling procedure. The published scores are therefore averages over runs that include outright deviations, not a curated subset.

What the paper does not contain is an independent benchmark against a human expert working the same tasks under the same conditions. The scores reflect expert assessment of the agent outputs, and readers weighing adoption should treat them accordingly.

What does this change for R&D teams?

The immediate value proposition is division of labor. If agents absorb software syntax and tool-switching, researchers can concentrate on physical assumptions and interpretation — the part of the workflow where judgement carries the most scientific weight. For labs with thin computational support, that shifts effort away from scripting and toward problem formulation.

The authors are explicit about scope. They present the agents as an aid to scientific judgement, not a replacement for it. The observed failure modes — a wrong observable, a mis-sampled initial state — are exactly the kind of silent errors that make unsupervised use risky, since an agent can produce plausible code and figures around a mistaken setup.

The practical near-term use is a supervised one: a researcher specifies the physics, the agents generate the code and results, and the researcher checks observables and sampling choices before trusting the numbers. The paper's full results are available at IOPscience.