What they did
Geby Jaff, writing on the Vals AI blog, set a team of Claude Opus 5.5 agents the job of finding Luttinger-compensated magnetic semiconductors: materials with zero net magnetism whose electrons are still sorted by spin, which spintronics memory research wants. The agents ran density functional theory calculations in Quantum ESPRESSO on cloud machines (Modal), at a fast PBE+U level and a slower HSE06 level, across 868 pw.x runs. They designed a new five-element oxide, YBaMnFeO5, simulated its band gap, spin windows and ordering temperature, and modelled whether its Mn/Fe checkerboard would survive synthesis. They also searched the literature and picked out KV[Cr(CN)6], a compound first made in 1999, as already having the wanted electronic structure. The author published every input, raw output and script in a public ledger with a one-command checker.
How it went
Two candidates came out. The designed oxide looks nearly ideal on paper (2.35 eV gap, ordering predicted around 420 to 490 K) but the agents' own cation-ordering simulation found the checkerboard scrambles at about 950 K, below normal synthesis temperatures, so it may not be makeable. KV[Cr(CN)6] is predicted to have a 2.1 eV gap with spin-sorted band edges and was measured in 1999 to stay ordered to 376 K; nothing about its spin sorting has been measured yet. Building the ledger exposed five errors in the agents' records: a hull-distance figure computed before competing phases had finished, an HSE06 run marked converged that was not, partly relaxed geometries, one truncated output, and a missed 2008 paper that already showed the spin-sorted edges, which narrowed the novelty claim.
Worth knowing
Nothing has been synthesised or measured; these are simulation results, and the agents' own adversarial review graded the designed material a design study and the 1999 material an identification with numbers rather than a breakthrough. HN commenters were sceptical, noting that undergraduates run Quantum ESPRESSO and that the expensive part is making and testing the crystal. The repository's corrections section is the honest part of the story: the agents' bookkeeping needed a human re-read.