Bringing quantum algorithms to practical use takes better hardware and better algorithms, but also better implementations of the algorithms themselves. This paper presents a system that puts AI agents to work on hard compilation problems in quantum computing, with every result certified by an exact verifier. An AI agent builds each verifier with a Phasecraft researcher. The researcher signs it off only after it has passed a set of validation tests, and it is then frozen before any solution is submitted.

Tested on three problems in quantum simulation, the agents beat the best published or standard methods on all three: up to 23% lower Pauli weight than a state-of-the-art optimizer for fermion-to-qubit encodings, up to 90% fewer SWAP gates for qubit layouts, and up to 73% fewer CNOTs in circuit synthesis. Every solution is a program that can be read, and the paper identifies the idea behind each improvement. The same approach applies to many more problems in quantum computing.