From Wet-Lab Feedback to Therapeutic Intervention Rules: GALILEO, a Recursive Embodied AI Scientist

Agentic systems for biology can reason over literature and propose experiments, while self-driving laboratories can automate selected assays. The missing capability is to convert physical observations into revised hypotheses, candidate-generation policies, and experimental tests over long discovery horizons. We present GALILEO, an embodied multi-agent system that closes this learning loop for therapeutic peptide discovery. GALILEO couples multimodal target nomination, sequence design, specialist computational tools, mixed robotic/researcher execution, and persistent Observation–Thought–Action–Summary (OTAS) memory. Starting from a clinically informed prior of 4,822 peptides, a frozen computation–diversity–agent funnel selected 10 first-round candidates. Across two five-round campaign records, feedback reorganized search from local sequence edits to scaffold and formulation exploration, while recurrent sequence–property relations were distilled into an Amphiphilic Balance Grammar (ABG). Crucially, the loop updates three scientific objects: the active mechanistic hypothesis, the policy used to generate the next intervention, and the verifier required for advancement. In the LRRC8C campaign, this produced cell-context engagement, direct taurine-transport, genetic, and membrane-integrity tests. In the mitochondrial SLC25A1 campaign, it produced intact-organelle localization, direct purified-target binding, citrate-transport, and metabolic validation. System-level benchmarks probe information coordination and sequential evidence use, whereas the two physical campaigns establish biological validity. GALILEO’s distinctive contribution is therefore not a single model or molecule, but an agent–experiment–learning architecture in which wet-lab evidence becomes persistent scientific memory and transferable intervention knowledge.