An Illinois-led preprint connects a scientific multi-agent planner to distributed robotic biofoundries. iBioGenie turns a natural-language research objective into hypotheses and protocols using literature, databases, and computational tools; BiofoundryAI compiles the plan into instrument workflows, schedules them, collects results, and returns the measurements for another design cycle. The important increment is physical closure of the loop—not another paper-writing agent or benchmark.
The three demonstrations are unusually concrete. In a phytase-engineering case, 64 of 75 proposed variants reportedly beat the previous best after four earlier optimization rounds had begun to plateau; the top variant improved 5.28-fold on that prior best and 70.17-fold on wild type. In a PAND enzyme search for yeast 3-hydroxypropionic-acid production, 18 of 28 successfully transformed candidates produced detectable product; a second mutation round ultimately reached 2.82-fold over the original starting enzyme. In a beta-carotene case, two screens tested 74 guide RNAs, yielding 21 effective guides across all 14 nominated knockout targets; the best double knockout improved production 1.69-fold over the single-knockout start.
The limitations keep this from becoming an autonomous-science victory lap. It is a non-peer-reviewed preprint with three synthetic-biology cases. The system remains domain-specific; large-plan scaling, token and GPU efficiency, prompt injection, privacy, and biosafety remain open. Its digital twin checks workflows before execution but does not yet learn from physical failures, and the full code stack is not yet generally released. Still, this is the right research-agent milestone to watch: whether an agent can carry a hypothesis through wet-lab execution, ingest the result, and make the next experiment measurably better. Preprint · Illinois biofoundry · Detailed figure review