Opus 5.5 Agents Found a Room-Temperature Magnetic Semiconductor Hiding Since 1999

Vals AI published a result on October 4 that doubles as a demonstration of what agent swarms can do to a scientific bottleneck: two candidate room-temperature antiferromagnetic semiconductors, found by a team of Claude Opus 5.5 agents running density functional theory simulations, with every input file, raw output, analysis script, and known caveat published on GitHub. One candidate is a newly designed five-element compound; the other was first synthesized in 1999 and, in the post’s framing, had been “hiding in plain sight” for a quarter century.

Why Room-Temperature Antiferromagnetic Semiconductors Are the Prize

The physics target is specific. Ordinary antiferromagnets cancel stray magnetic fields, which allows dense packing and roughly 1,000-times-faster switching than ferromagnets, but they cannot sort electrons by spin, which spintronics needs. The sought material is “Luttinger compensated”: zero net magnetism overall, yet with spin-up and spin-down atoms on inequivalent sites so spins are separated by energy at the band edges. The catch is thermal jitter: at room temperature, that jitter is about 26 meV, so a usable material needs a spin-sorted window at the band edges vastly larger than that. A 2025 study called a room-temperature Luttinger-compensated semiconductor “the open goal” in the field. The second candidate may already fill it.

The Two Candidates, and the Honesty About Each

The first candidate, YBaMnFeO5, is newly designed: a predicted 2.35 eV band gap with spin windows of 1.0 eV (holes) and 1.4 eV (electrons), stable to roughly 420 K. The caveat is serious: it requires a perfect manganese-iron checkerboard, and disordering simulations show that collapsing around 950 K, while synthesis happens at 900 to 1,300 degrees Celsius, meaning making it might destroy the property that makes it interesting. The second candidate, KV[Cr(CN)6], is the remarkable one: a Prussian-blue-family compound made in 1999, whose zero net moment was intentional and whose spin sorting was visible in a 2008 paper, but which nobody had labeled a Luttinger-compensated semiconductor or quantified its windows. The agents’ analysis found roughly a 2.1 eV band gap with spin windows of 2.6 and 1.6 eV, and the 1999 sample stayed magnetically ordered to 376 K, above room temperature. Its structure locks the metals in place, solving the checkerboard problem by chemistry rather than luck. The post’s conclusion: “materials for the next step in spintronics may already exist, waiting to be recognized.”

The Workflow Is the Reusable Part

The methodological details are what make this repeatable rather than a one-off. The agents ran DFT simulations at two levels of approximation, a fast PBE+U screen and a slower, more accurate HSE06 confirmation, with band gaps and spin windows reported from HSE06 only. Everything shipped: input files, raw outputs, analysis code, a one-command checker, independent re-runs, and a disclosed list of caveats, including that methods disagree on whether water in the 1999 sample preserves or halves the spin window, and that nothing has been experimentally measured yet. Compare the rigor posture with the Claude enzyme-discovery post: that result had lab validation but the analysis pipeline stayed with Anthropic; this result is computational-only but fully reproducible by anyone with the GitHub ledger. The two together sketch the emerging standard for AI-for-science claims: agents propose, humans or experiments validate, and the ledger is public either way.

What to Watch

Three things. First, the experiment that matters: remaking KV[Cr(CN)6] and directly measuring its spin sorting, which any enough-equipped lab can do from the published data. Second, whether the materials-science community adopts the two-tier DFT screen-and-confirm workflow as a standard for computational claims, since the agents’ value here was running thousands of accurate simulations that no single lab would prioritize. Third, the meta-observation: the most consequential find was not the new compound but the 1999 one, which suggests agent swarms searching existing literature and samples for unrecognized properties may yield more near-term wins than de novo design.