Self‑driving laboratories (SDLs) combine automated experimentation with adaptive decision‑making to speed up scientific discovery, yet they usually rely on specialists to translate research goals into closed‑loop campaigns and to adjust them as data or conditions change. This work introduces La Agente Optima, an agentic framework that builds and supervises Bayesian‑optimization campaigns across computational and experimental platforms while preserving a persistent optimization state. Optima separates large language model (LLM) reasoning from the executed loops, running repetitive optimization steps autonomously and handing control back to the agent only when interpretation or campaign revision is required, thus keeping every decision auditable. We evaluated Optima through ablation studies, five digital discovery tasks, and two physical platforms. In a closed‑loop contact‑angle optimization, Optima detected and corrected a mid‑run measurement failure, improving the angle from 71.4° to 67.8° (just above the 64‑66° target range), and inferred that the goal was likely unattainable with the available reagents, recommending a formulation change. In a five‑day multi‑objective flow‑chemistry campaign, Optima raised the yield from 30% to 59% over 23 experiments; despite higher inference costs, it used substantially less starting material and incurred lower overall cost than a human‑directed campaign, while selecting a more mass‑efficient operating point. These results show that LLM‑based agents can conduct rigorous, long‑running optimization without specialist setup, expanding the scope of SDLs.
Review