What five agents actually do to a clinical intake file

A clinical intake file arrives as whatever the patient's history happens to be: scans, letters, lab results, none of it structured. Turning that into a protocol a clinician will sign used to take four days.

The pipeline that does it now is not complicated. Five agents, each with one job. One parses the intake file and pulls structured values out of documents that were never structured. One assesses risk against those values. One drafts the protocol. One handles lifestyle guidance. One assembles the report a clinician actually reads. They run on Azure OpenAI inside Swiss and EU cloud isolation, because the alternative was not an option for this client.

Splitting the work that way is not about speed. It is about being able to answer the question a reviewer will ask, which is never "how fast" and always "where did this come from". A single model doing all five jobs gives you an answer and no trail. Five agents with defined handoffs give you a chain you can walk backwards, which is why every recommendation carries its citation back to source.

That is the part that took the time. Parsing biomarkers is a solved problem. Producing output a clinician will put their name to is not, and the difference is almost entirely in what happens after generation rather than during it.

The honest version of the four-days-to-five-minutes claim is this: the five minutes is the generation. The reason it can be trusted in five minutes is the evaluation layer around it, which we built first and which never appears in the headline.

interesting" beat. That belongs to "The number in the case study is never the interesting part", per the shared-facts rule.

are public on `/pipeline` with the related-party disclosure. Remaining blockers are Pierre writing this, agreeing the credit, and a verifying URL.