On October 16, Google presented three research components connecting agents to physical experiments. VERID represents the laboratory through a persistent semantic twin, supporting inspection, pre-execution simulation, traces and execution checks. Its report validates optical and fluidic foundations using dye and water mixtures, not universal automation of biology.
A second study uses three design, test and improvement cycles for aptamers against alpha-thrombin and myoglobin. Experimental measurements help revise strategies and nonspecific binding; preliminary nanomolar affinities concern literature-derived motifs reworked by agents, whereas strictly de novo scaffolds show micromolar binding. Conflating these categories would inflate the result.
The third report studies epitope excision to reduce compute required for protein-binder design. It compares different autonomy levels and finds the human-agent workflow produces the most high-confidence designs per GPU-hour in its experiment. Yeast-display validation is neither an approved medicine nor clinical evidence.
The date belongs to the research post; no verified first-publication date is established here for the linked PDFs. Results are limited to the studied protocols.