On October 16, Google Research published a post on the building blocks for autonomous wet labs. The first of the three projects it describes is VERID, detailed in a companion paper by the same team. VERID sits between an AI agent's plans and the liquid-handling robot that carries them out.

It is serious work. Underneath the robotics, it makes an argument I have been making for years from the other side of the bench.

What VERID actually does

VERID (VERified Infrastructure for Discovery Laboratories) is a persistent digital twin of a robotic workcell. It tracks the position and geometry of everything on the deck, from robotic arms and plate readers down to individual wells and pipette tips, along with the identity, volume, and liquid height of every reagent. It is built on the open-source PyLabRobot framework and provides four core services:

VERID's four core services

  1. Introspection. The agent can query the live workcell, for example how much liquid is left in a given well, before it acts.
  2. Simulation. Every block of commands first runs against a virtual copy of the deck, fast-forwarding through incubations, before any reagent is touched.
  3. Trace persistence. Every run produces a dossier of state snapshots, camera video, instrument telemetry, and a line-by-line log linking protocol code to hardware commands.
  4. Safe execution. Unsafe or destructive commands are filtered out before they reach the hardware.

The failures lived in the execution

To keep biological complexity out of the picture, the team ran every case study with food coloring and deionized water on a Hamilton liquid handler. Even then, execution was where things went wrong.

In the first study, an agent ran a serial dilution and found each step diluting only 1.4 to 1.8 fold instead of 2. It guessed the mixing was insufficient. A human operator then reviewed synchronized side-view video against the execution trace and found the real cause: pipetting at the very bottom of the well left dense dye stratified there, so roughly 68 to 70 percent was carried forward at each step instead of 50, and some tips pressed flush against the well floor. Once the agent added a 1.5 mm clearance and changed how it mixed, the dilutions came back to 1.92 to 1.99 fold.

In the second study, the simulator rejected a plan to lower all eight pipette channels into a reagent trough, because the tips span at least 72.0 mm and the trough is 71.2 mm long. The agent switched to mixing with seven channels. After the run, Channel 8 appeared to read 28 to 47 percent low across all dyes. The agent designed a factorial diagnostic and found that Channel 8 was the only channel behaving correctly: the trough mix had coated the inner walls of the other seven tips with concentrated dye, adding about 4.9 µL (40.6 percent) of carryover.

As the paper notes, simple replication could have fixed neither error.

A property of physical science

Those findings could have come out of my notebooks. I spent over 10 years building genomics tools at QIAGEN, Bio-Rad, and Fluent BioSciences, which I co-founded and which Illumina acquired in 2024. The lesson I kept relearning is that the hardest part is rarely the invention. It is execution in the real world.

The authors put it plainly: conventional automation logs only whether a script completed, discarding the physical details of how each well was handled. VERID records those details, including channel assignments, tip reuse, aspiration heights, and dwell times, and turns them into data an agent can query. If a liquid handler with a live digital twin still needs that record, the gap between the protocol as written and the protocol as run is not a human failing. It is a property of physical science.

Most runs still happen at a bench

VERID instruments the robot, and that is the right place to start for closed-loop discovery at scale. The paper is also candid about the bench, where researchers can customize protocols but the work is labor-intensive and prone to operator-to-operator variability.

Those manual runs have no digital twin, no simulation gate, and no trace dossier. When something goes wrong, the scientist troubleshoots alone, and the kit's maker hears about it weeks later, if at all.

Automation does not erase the human role. In the first case study, it was a person watching the video who found the root cause, and the paper describes scientists annotating each run, "systematically capturing tacit laboratory know-how in machine-readable form." When I toured Ginkgo's autonomous labs during REAL 2026 in September, three people were needed every day just to prepare the instruments. The human work had moved, not disappeared.

Capture execution before you automate it

New methods are developed by hand, refined over dozens of failed runs, and only then handed to a robot. If those early runs go unrecorded, the agent inherits a protocol stripped of exactly the details VERID's case studies show matter most.

The bench deserves the same kind of record we are now building for robots: what was done, in what order, for how long, and where it departed from the instructions. That is why I proposed an open standard for protocol execution records in September. Google's results make the case for it more clearly than I could.

Full disclosure: I founded Adept Scientific, which makes Kavia, a free voice copilot that guides scientists through protocols at the bench and records how each run actually went.

Sources

  1. Google Research, Grounding AI agents in biological reality: Building blocks for autonomous wet labs (October 16, 2026)
  2. Delley et al., Semantic Representation for Agentic Self-Play in Automated Laboratories (Google Research, Google DeepMind, Google Cloud, October 22, 2026)