Scenario

An optimization with many interacting variables is too large for a fixed DOE and too slow for one-at-a-time trials. A self-driving campaign plans the space, lets active learning pick the most informative next experiments, dispatches them to the robot, analyzes the results, and repeats — converging with far fewer runs.

Worked example. A reaction yield must be maximized across temperature, catalyst loading, and solvent ratio within a fixed experiment budget.

Agent workflow

1. Experiment campaign planner — define the variables, objectives, and budget; seed an initial design.
2. Active-learning loop — fit a surrogate model and propose the next batch of experiments by acquisition value.
3. Robotic protocol dispatch — translate the batch into robot instructions and run it.
4. Lab result analysis — parse the measured results, update the model, and decide whether to continue or stop.

Demo output

A convergence trace showing the objective improving batch over batch, the proposed-vs-measured values, and the stopping decision. The magic moment: the loop reaches the optimum in a fraction of the runs a full-factorial DOE would need.

Deterministic vs LLM

• Deterministic — the DOE planning, Bayesian optimization, and result analysis.
• LLM — the campaign-strategy narrative and the stop-decision reasoning.

Every proposal and update is a validated calculation. The agent reads and routes; the engine decides.

Limits

Requires a lab that can execute the dispatched protocols; screening-level optimization, not a guarantee of the global optimum. It drives the search; the lab produces the truth.