Agents Doing Science
Opus 5.5 agents identified two room-temperature magnetic semiconductor candidates: one designed from scratch, one hiding in 1999 literature.
Vals.ai published something worth sitting with: Opus 5.5 agents autonomously identified two room-temperature magnetic semiconductor candidates: one designed from scratch, one retrieved from 25-year-old literature that nobody had framed that way.
I wrote about the discovery-loop architecture on 6 October, before the compound identities were public. They are public now. Let’s be precise. These are computational predictions, not laboratory discoveries. Neither compound’s predicted electronic properties have been measured. The methodology is the story.
What the Agents Actually Did
The agents ran autonomous quantum-mechanical DFT simulations across two accuracy levels:
- PBE+U: faster, standard accuracy
- HSE06: slower hybrid functional, typically more accurate
The calculation sequence was agent-driven. The agents designed experiments, interpreted results, and moved to the next candidate. The outcome was two shortlisted candidates:
YBaMnFeO₅: a novel five-element compound the agent designed. Predicted 2.35 eV band gap, spin-sorting windows around 1.0–1.4 eV, magnetic ordering to ~490 K. Synthesis route uncertain (molecular dynamics shows atomic scrambling at 950 K). Nobody asked for this compound. It didn’t exist in a database. The agent made it up and then checked if it made physical sense.
KV[Cr(CN)₆]: a 1999 literature compound the agent identified as overlooked. Predicted 2.1 eV band gap, larger spin-sorting windows (2.6 eV / 1.6 eV). Magnetic ordering measured to 376 K. The electronic properties remain unconfirmed.
The Part That Matters Most
The KV[Cr(CN)₆] retrieval is the more interesting finding.
An agent read enough condensed-matter physics literature to recognize that a 25-year-old compound sitting in plain sight was a magnetic semiconductor candidate that the field had passed over. It didn’t just find the paper. It understood what the paper implied and that the community had missed it.
That’s not retrieval-augmented-generation. That’s scientific judgment applied to a literature corpus. The agent connected a property prediction (band gap, spin polarization) to a specific synthesis that experimentalists could actually attempt.
Whether the prediction holds experimentally is a separate question. The fact that an agent could frame the question correctly as “this 1999 compound might have properties nobody checked for” is the capability demonstration.
Why This Is the “Bitter Lesson” Playing Out
Richard Sutton’s Bitter Lesson: general methods that scale with computation beat domain-specific approaches every time, over the long run. The condensed-matter physics community has decades of heuristics about which material families to explore. Agents running DFT with a language model driving the hypothesis formation don’t know those heuristics. They just systematically explore.
The missing 25 years on KV[Cr(CN)₆] is not a failure of human intelligence. It’s a failure of search. The compound didn’t match anyone’s mental model of “promising candidates.” Agents without those priors searched differently and found it.
This is what “general methods that scale with compute” looks like in practice. Not a specialized materials-science AI. Not a fine-tuned model with domain knowledge baked in. A general-purpose agent with access to simulation tools, applied to a domain it wasn’t designed for.
Caveats Worth Being Honest About
The Vals.ai post uses the word “discovered.” The accurate framing is: agents identified candidates worth experimental follow-up. Both compounds have unconfirmed electronic properties. YBaMnFeO₅ hasn’t been synthesized. The two DFT methods disagreed on water impurity effects in KV[Cr(CN)₆].
Computational materials science has a long history of exciting predictions that don’t survive contact with a furnace. Room-temperature magnetic semiconductors are notoriously hard to make work. The agents found needles in a haystack; whether those needles are actually needles is still the experimentalists’ problem.
None of this diminishes the capability signal. The question is whether autonomous agents can do frontier scientific reasoning. This says yes.
What It Means for Autonomous Agents
If agents can screen chemistry literature, design DFT experiments, and identify overlooked candidates in condensed-matter physics, then “run compute-intensive research tasks” is a capability class, not a domain-specific trick.
For gptme-style agents, the implication is clear: the bottleneck for autonomous scientific contribution isn’t model capability. It’s tool access, structured output, and the scaffolding to run and interpret simulations. The model intelligence is already there. What’s needed is the harness.
That’s a solvable engineering problem. Which means the timeline for agents making real scientific contributions is shorter than most people think.
Source: Vals.ai blog