La Agente Óptima made self-driving lab campaigns auditable and cheaper

Self-driving laboratories can run experiments adaptively, but specialists still have to translate goals into executable closed-loop campaigns and adjust them as conditions change.

The team built La Agente Óptima around a persistent optimization state. It constructs and supervises Bayesian-optimization campaigns across computational and physical systems, while separating LLM reasoning from the campaign actually being executed. Routine loops run consistently; control returns to the agent when progress needs interpretation or revision.

In a contact-angle campaign, it caught a mid-run measurement failure, moved the result from 71.4 to 67.8 degrees, and inferred that the 64–66 degree target was probably unattainable with the available reagents. In a five-day flow-chemistry campaign, 23 experiments lifted yield from 30% to 59%; the campaign cost less and used less starting material than a human-directed one.

Separate LLM reasoning from campaign execution

Persistent state stops each loop from resetting its context.

Separating execution from reasoning makes repetitive optimization consistent.

Auditable decisions make measurement failures and campaign changes diagnosable.

The agent can recognize an unattainable target and recommend a formulation change instead of iterating blindly.

For long-running optimization, separate reasoning from execution: let the system repeat validated loops, but reserve human or agent attention for interpretation, failure recovery, and reformulation.

The paper reports five digital discovery tasks and two physical platforms, with executable campaigns maintained as the scientific problem and execution environment evolved. It presents Óptima as a way to make rigorous, long-running optimization accessible to domain scientists without specialist setup.