An Agent Runs Bayesian Optimization Campaigns and Only Interrupts the LLM When Interpretation Is Needed

arXiv 2609.04564·low signalXBlueskyLinkedInCopy link

arXiv 2609.04564 presents La Agente Optima, which constructs and supervises Bayesian optimization campaigns across computational and experimental systems while holding a persistent optimization state. The design choice worth stealing is separating LLM reasoning from executed campaigns, so repetitive optimization loops run consistently and control returns to the agent only when progress requires interpretation or campaign revision, keeping every decision auditable. It was evaluated across ablation studies, five digital discovery tasks and two physical self-driving-laboratory platforms.