EmulatRx: five agents designing a clinical trial

A few days ago I came across a paper that caught my attention: EmulatRx, published in Nature Communications.

It’s a multi-agent system that designs clinical trials from real-world data, coordinating five specialised agents that build the protocol, map it to the electronic health record and estimate the treatment effect.

It’s a good example of how this architecture will make its way into clinical research: task by task, not as a single product.

Five things that match what I see in practice:

• Real-world data is still the hard part. Missing data, heterogeneity across sources, protocol variables that simply don’t exist in the record. These are the real problems you hit when working with real data.

• One profile per agent. Trialist, informatician, clinician, statistician. They interact and they decide: when the statistician spots a covariate imbalance, it asks the clinician.

• Traceability of decisions. You can reconstruct why the system handed off from one agent to another. In a regulated setting, that’s what decides whether you can use it or only demo it.

• Integrated knowledge sources. ClinicalTrials.gov and PubMed. The model doesn’t recall: it looks things up.

• Evaluation and optimisation still depend on human expertise. There’s no shortcut. The system speeds up the cycle, it doesn’t replace the judgement.

None of this is specific to trial design. Messy data, specialised roles, traceable decisions, grounded knowledge, a human at the end: that’s the shape of anything that has to survive in a regulated environment.