Google Research posted something wild this week, and it’s the clearest example yet of agents leaving the screen and touching the physical world. They’re hooking AI agents up to real lab robots so the agents can design molecules, run the experiments, read the results and try again.
The core problem: an agent can write code that’s valid but physically impossible, like lowering a pipette into a trough that’s too short, or drawing from an empty well. Their fix is a framework called VERID, a live 3D “digital twin” of the lab that tracks robot geometry, well volumes, reagents and pipette tips. The agent plans and simulates every step before the robot touches anything.
With that in place they ran closed-loop design cycles and say they found single-digit nanomolar protein binders for cancer targets.
This is the same idea I keep coming back to with agents at home: give them real guardrails and a true picture of the world, and they get a lot more useful. Same lesson, much bigger stakes.
Happy Hacking!
