Agentic Lab: An Agentic-physical AI system for cell and organoid experimentation and manufacturing

Abstract Reproducibility in biological research and manufacturing remains constrained by the complexity of multi-step protocols, fragmented data-analysis pipelines, and the intrinsic variability of experimental execution. Here, we present Agentic Lab, an agentic-physical AI platform that unifies large language model and vision language model (LLMs/VLMs)-driven reasoning with real-world laboratory operations. Agentic Lab uses multi-agent orchestration architecture, comprising of specialized subagents for knowledge retrieval, protocol design, multimodal data analysis, and training-free segmentation and representation learning for intrinsically explainable single-cell and organoid phenotyping. These agents operate under the orchestration of a virtual principal investigator MolAgent that is linked to an augmented reality (AR)-based physical AI interface, which can bridge digital reasoning with human physical execution. Agentic Lab perceives real-world experimental activities, provides context-aware instructions, identifying procedural errors in real time for humans to correct, and continuously evolves with its long-term memory database expanding through the accumulation of experimental data logs from human scientists. This interaction allows scientists and AI agents to collaborate and co-evolve dynamically, closing the loop between planning, action, and analysis in the traditional cell and organoid research lifecycle. We demonstrate Agentic Lab in organoid differentiation from human pluripotent stem cells, where it autonomously generates protocols, monitors culture procedures, and identifies subtle morphological heterogeneity linked to growth conditions. The system interprets these phenotypes, grounds them in literature, and proposes targeted instructions for improving differentiation efficiency. By combining multi-agent reasoning with physical laboratory awareness, Agentic Lab transforms experimentation and biomanufacturing from a static workflow into an adaptive, feedback-driven, bidirectional process that integrates agentic AI into the research lifecycle. This framework establishes a foundation for intelligent laboratories that integrate design, execution, and interpretation within a unified agentic-physical system.