Screen new magnetic materials for computer memory with an agent team

A researcher used a team of Claude Opus agents to run physics simulations and surface two candidate magnets for spin-based memory, then published every calculation.

By Geby Jaff (Vals AI) on Vals AI blog · Posted Oct 4, 2026 · Added Oct 15, 2026 · 2 sources

Difficulty: AdvancedOur rating
Time to build: Not reportedOur estimate
Monthly cost: Not reported
Revenue: Not reportedWe only show revenue the source states

What they did

Geby Jaff (Vals AI) directed Claude Opus agents that ran density functional theory simulations to look for semiconductors with zero net magnetism that still sort electrons by spin. The agents designed one new compound and flagged a second material first made in 1999.

How it went

The post reports two candidates: a designed compound that may be hard to synthesize in its useful form, and KV[Cr(CN)6], whose 1999 sample stayed magnetically ordered up to 376 K. Neither the band gap nor the spin sorting has been measured yet. Inputs, outputs, code and known caveats are in a public GitHub repo.

Worth knowing

These are predictions, not lab results. The two simulation methods disagree on how much water in the real sample weakens the effect. Shipping a public ledger with a one-command checker is what makes the claim reviewable.

Builder recipe

Our own starting point, written from the source. Adapt it to your setup.

Stack

Claude (Opus-class) agents
Simulation software your field already trusts
A public repo with inputs, outputs and a checker script

Steps

Write the target properties as hard numeric criteria.
Let agents propose candidates and run the cheap simulation first, the expensive one on survivors.
Have a separate agent list every caveat and failure mode.
Publish all inputs, raw outputs and a one-command re-run so others can check.

Starter prompt

Our target: [property list with numeric thresholds]. Propose 20 candidates, run the fast screen on each, and keep only those that pass every threshold. For survivors, run the accurate method and write a caveats section listing every way the result could be wrong.