Can 90 AI agents discover a new material in three days?

Not a giant lab. A small outside team, over 90 agents built on Claude Opus 5.5, and three days on the clock. The target: a rare kind of magnet, called Luttinger compensated, that cancels its own net magnetism yet still sorts electrons by spin. Physicists have been dreaming about one for decades.

The result? Two candidates. One is a brand new compound. The other has been sitting in the literature since 1999, and nobody had noticed it fits.

Now the catch. It is all computation. Nobody has made either material in a lab yet.

Why would anyone care?

Because of spintronics: storing and moving information with the spin of an electron instead of its charge. The promise is memory that is faster, less power hungry and longer lasting than what we have now. The other thread is quantum computing, where fragile quantum states need protection from noise. And a magnet with no net field is handy, because it does not disturb its neighbours when you pack parts tightly.

So which is the breakthrough: predicting a material, or actually building it?

Building it is the hard part. Plenty of materials that look great on paper cannot be made, or lose their magic in the real world. On the other hand, the data, scripts and checking tools were published, so anyone can poke at the claims and try to break them.

And that may be the bigger story. Discovery is no longer only for the giant AI labs. A small team with a swarm of agents can just go and try.