Lab automation just got a layer of AI oversight — and a real-world test on actual protein assays.
Researchers published a framework that uses two AI agents to bridge the gap between how scientists write experiment protocols (plain English) and how lab robots actually operate (rigid control commands). A Parser Agent converts the natural-language protocol into a structured format; a rule-based engine then maps that structure onto device-level commands for the robotic platform. A second, independently chosen model — the Validation Agent — checks the output for completeness, parameter accuracy, and correct execution order. When it finds errors, it triggers a self-correction loop. The team ran a sweep across 7 Parser models and 3 Validator models on ELISA protocols to measure how model scale and validator choice affect accuracy and pass rates.
Microplate experiments are a useful stress test because they require juggling well mapping, sample-reagent combinations, replicate placement, and parallel dispensing all at once — exactly the kind of multi-constraint problem where a single LLM piping output straight to hardware tends to fail silently. Cross-model verification, where the validator is a different model than the parser, adds an adversarial check that a self-reviewing single agent can't provide.
The system was validated end-to-end on a Bradford assay — a real protein quantification experiment, not a simulation — which is a more credible proof point than most lab-automation papers reach. Whether it scales beyond ELISA and Bradford assay protocols to the messier long-tail of bench science remains the open question.