What could an agentic laboratory look like?

An agent connects the tools

Start with the compound and the evidence

Retrieve the structure and available experimental measurements, with scientific context from laboratory procedures.

Conceptual view of the published APPA workflow. Select a stage to follow the information flow. This is an explanation, not a running agent.

The scientific question

Preformulation draws on measurements, models and laboratory knowledge. My broader question is how an agentic ecosystem can connect those resources into useful scientific workflows.

The published starting point

APPA joins an LLM with databases, predictive models and empirical calculators. The paper focuses on Developability Classification System assignments and preformulation guidance.

What the study establishes

A 500-case virtual-compound exercise tests tool use against a context-provided GPT-4o baseline. Four marketed-drug examples also match their reference DCS classes. These are workflow evaluations, not evidence of autonomous laboratory operation.

What comes next

I am working on follow-on papers about agentic workflows and ecosystems for future laboratories. The published APPA study identifies missing-input handling as a weakness: an agent may invent a value instead of asking. Reliable coordination needs to preserve the distinction between measurements, predictions and missing information.