La Agente 'Optima' introduces an agentic framework designed for self-driving laboratories (SDLs). It constructs and supervises Bayesian optimization campaigns across both computational and experimental systems, maintaining a persistent optimization state.
The framework separates large language model (LLM) reasoning from campaign execution. This allows 'Optima' to run repetitive optimization loops consistently, only involving the agent when interpretation or campaign revision is necessary. Every decision remains auditable.
Evaluation of 'Optima' included ablation studies, digital discovery tasks, and physical platforms. It successfully adapted campaigns as scientific problems and environments evolved, such as correcting measurement failures and optimizing chemical yields.
Results indicate that 'Optima' can perform rigorous, long-term optimization campaigns with less resource use than human-directed efforts, making advanced optimization accessible to domain scientists without specialized setup.
Source: https://arxiv.org/abs/2609.04564