Health AI Central

The problem

Health-AI information is scattered: research literature, trial registries, regulatory databases, company announcements, and specialist reporting each tell part of a story, and nothing connects the story to the evidence. I wanted a publication that reads like specialist journalism and cites like a systematic review, and I wanted it to run every day without a newsroom.

The design

More than forty agents do the daily editorial, research, and promotion work; the monitoring registry tracks 66 scheduled jobs in all, most on a staggered overnight cron. The model split follows one rule: Sonnet for judgment-heavy steps, Haiku for high-volume steps. Classifying a thousand headlines is volume; deciding what deserves 1,500 words is judgment.

The map below shows one slice, the editorial pipeline, the most visible of the agent families:

A night runs like this. Ingestion agents pull from news lanes covering 1,000+ companies, 25 RSS feeds, and primary sources (PubMed, ClinicalTrials.gov, FDA databases, NIH RePORTER), most without any model at all. A story listener (Sonnet) scores what came in; the content manager sets the day’s editorial brief; the medical researcher picks 10–12 stories and fetches each source’s full text; the technical writer (Sonnet) drafts news briefs of 400–700 words and deep dives of 1,200–2,000; the editor (Sonnet) reviews and calls the visual artist, which writes an image brief in Sonnet, renders it with Gemini 3 Pro Image, and then scores its own output.

Two other families work alongside the editorial pipeline. Research-intelligence agents screen the incoming data, link records across databases (a trial registration to its sponsor, its publications, and its regulatory record), and generate the intelligence products users come for: company profiles, clinical-area briefs, and classifications of why trials stopped. Back-office agents run the business itself: customer service and inbox triage, traffic analytics, SEO, security monitoring, advertiser operations, and the site's own help assistant. The same rules apply to every family: one gateway, per-agent keys, and a monitor that watches every job.

Hand-offs are files and database states, not vibes. The pipeline is a state machine with an explicit list of allowed transitions; an illegal transition raises an error instead of publishing a mistake. Every model call goes through one gateway with a per-agent key, so each agent’s spend and behavior are individually visible.

Evaluation and guardrails:

• Every article must link a primary record: a DOI, PMID, trial number, FDA clearance number, or NIH project number.
• A source-fidelity reviewer (Sonnet) checks the draft against its sources; a duplicate-story guard checks it against everything already published.
• AI-generated images must score 80 or better in review; source images must pass a provenance check.
• At most two automated revisions, then a human hold.
• For research classification, the evaluation went further: a blind three-vendor judge panel decided disagreements by strict majority, and clinicians checked a subsample. Sealed result: F1 0.953.

And the guardrail that matters most: nothing publishes without me. The pipeline’s last automated act is an email. I approve, reject, or hold every article at Board Review, and my rejections are fed back into the next day’s story selection as negative examples. I reject about a quarter of what reaches me, which is the system working, not failing.

The numbers

Figure	Value
Agents in the daily pipeline	~40 (66 scheduled jobs in the monitor registry)
Articles published since launch	500+
Drafted per day	~10–11
Published per day	5.3 → 6.9 (month over month, Oct 2026)
Quality-hold rate	34.7% → 14.5% after the contract change
My rejection rate at Board Review	~25%
Monthly model spend	≈ $200 via the gateway

Media

[TODO: 60–120 second screen recording of the pipeline triaging a day's health-AI news]