Google's VERID lets AI agents run wet-lab robots

Google Research has laid out three systems meant to tie AI agents to what is physically true inside a laboratory: VERID, a digital twin of the lab workcell; a closed-loop pipeline for designing DNA aptamers; and an epitope excision workflow for protein binders. The October 16, 2026 post is credited to Pooja Rao, Petar Sirkovic and Annalisa Pawlosky. All three attack one bottleneck — models now propose experiments faster than hands and robots can run them.

A digital twin standing between the agent and the hardware

VERID is short for VERified Infrastructure for Discovery Laboratories. At its center sits a semantic digital twin: a live, stateful 3D tree that tracks robotic arm geometry, the volume in each individual well, reagent identities and pipette tip inventory. Google lists four capabilities — live 3D state tracking, virtual dry runs that finish in seconds, hardware safety checks that refuse invalid commands, and synchronized video and action logs. It runs on the open-source PyLabRobot framework.

The reason for that scaffolding is unflattering to the models themselves: frontier systems, Google writes, do not reason reliably about three-dimensional geometry or physical state. So no unverified code reaches the robots. Simulation is mandatory, and humans authorize sequences, protocols and runs.

Aptamers as the test bed

The team picked DNA aptamers binding myoglobin and alpha-thrombin as a tractable sandbox. Across three design-test-iterate cycles, myoglobin hit rates climbed from 10.2% to 33.5%, ending at 72.4% specificity. Alpha-thrombin produced no binders at all in round 1, reached 26.0% in round 2, then 42.1% at 98.5% specificity in round 3. Each round consumes 7 to 10 days of wet-lab time.

The dip in the middle is the story

Round 2 of the myoglobin campaign collapsed to a 3.7% hit rate and 19.1% specificity. The agents had over-optimized for G-quadruplex structures, and 63% of that round's designs turned out to be sticky, non-specific binders. Reading their own assay data, the agents capped the promiscuous motifs, subtracted bead binders and pivoted to non-G4 scaffolds — a ninefold recovery in on-target hit rates.

The alpha-thrombin shutout traced to fixed primer overhangs that disrupted folding in 48% of sequences. The agents added clamped flanking stems in later rounds. In both cases the correction is presented as self-diagnosed from experimental results rather than handed down by a human.

Binders against FAP

The third block targets fibroblast activation protein alpha, or FAP, a protein of interest in tumor research. Rather than predicting whole structures, the workflow excises the relevant epitope, which accelerated each cycle more than eightfold computationally. Human-agent collaboration nearly doubled in-silico hit rates against humans working alone. Candidates were validated with yeast surface display and fluorescence-activated cell sorting, yielding binders at single-digit nanomolar affinity.

Guardrails, and what Google says is unsolved

Experiments are confined to well-controlled, non-infectious biological systems with no gain-of-function risk, every designed sequence passes biosecurity screening, and the work is framed against Google's Frontier Safety Framework. The open problems are stated plainly: making sure agents optimize for genuine target binding instead of assay shortcuts, verifying multi-step wet-lab protocols, and continuously monitoring agent reasoning and execution traces.

Two things this article will not assert, because the post does not support them. No agent model is identified — there is no Gemini attribution in the text. And no cost figures are given for any campaign. Three accompanying papers are referenced; this write-up rests on the blog post alone, not on those papers.