Google Research Unveils VERID and Closed-Loop Agents for Autonomous Wet Labs

VERID provides four core capabilities:

• Live 3D state tracking: Lets agents verify deck layouts, available tips, and liquid volumes before running actions.
• Fast virtual dry runs: Simulates protocols in seconds to identify potential hardware collisions or attempts to aspirate from empty wells.
• Hardware safety checks: Automatically halts invalid commands before instructions reach the physical robot.
• Synchronized video and action logs: Records execution data to support debugging, reproducibility, and the separation of biological signals from physical noise.

In practical case studies, VERID prevented an agent from lowering an 8-channel pipette into a narrow trough that was too short, prompting the agent to replan on the fly and design a targeted diagnostic for tip-wall carryover. In another test, persistent state tracking allowed an agent to recover mid-run from a plate-reader failure without re-pipetting reagents. Furthermore, during a self-play campaign to reconstruct blinded dye mixtures using camera images, an agent adjusted for fluidic and optical artifacts to improve measurement accuracy by more than fivefold.

Closed-loop discovery for aptamers and proteins

Google Research validated its agentic approach across two wet-lab discovery domains: DNA aptamers and de novo protein binders.

In the aptamer project, agents conducted three design-test-iteration cycles on two targets: myoglobin and α-thrombin. Because DNA can be chemically synthesized in pooled libraries and read through sequencing, each cycle took just 7 to 10 days. When early designs hit laboratory hurdles, the agents adapted autonomously. On α-thrombin, Round 1 yielded 0% binders due to folding disruptions from fixed primer overhangs; after adding clamped flanking stems and diversifying sequences, hit rates reached 42.1% at 98.5% specificity in Round 3. On myoglobin, agents initially over-indexed on G-quadruplex motifs in Round 2, causing off-target binding to bare magnetic beads. After detecting the artifact in the readout data, the agents capped promiscuous motifs in Round 3, recovering specificity to 72.4% and achieving a 33.5% confirmed hit rate.

For protein design, the team targeted fibroblast activation protein alpha (FAP), a receptor overexpressed in solid tumors. Because re-folding complex proteins in silico scales quadratically with target length, the team implemented an agent-guided epitope excision workflow that excised only the local binding pocket during folding. This strategy accelerated compute cycles more than eightfold without sacrificing structural accuracy. Tested in the wet lab via yeast surface display and fluorescence-activated cell sorting (FACS), the workflow yielded binders with single-digit nanomolar affinity within a few runs. Under equal compute budgets, human-agent collaboration nearly doubled the in-silico hit rate compared to a scientist working alone.

What it means for developers

Google Research's findings highlight a shift in how developers can approach AI agents interacting with real-world environments. Rather than relying on LLMs to write raw, unconstrained execution scripts, practical agentic systems require intermediate validation layers. VERID demonstrates that digital twins, state tracking, and rapid dry-run simulations are necessary safeguards when autonomous systems interact with physical hardware.

For engineers building agentic pipelines—whether in life sciences, robotics, or backend automation—the research also shows that multi-agent systems perform best when given rich trace logs and multi-modal feedback to detect physical artifacts. As multi-model architectures become standard for complex reasoning tasks, developers can experiment with top AI models cheaply through one API at https://apixoai.online to evaluate how different models handle structured planning, code generation, and failure recovery.

Google noted that all workflows adhere to its Frontier Safety Framework, incorporating biosecurity screening, non-infectious test systems, simulation-first execution, and mandatory human authorization for protocols and sequence designs.

Source: Grounding AI agents in biological reality: Building blocks for autonomous wet labs — Google Research. Written by the Apixo team from that report.