Reaction to Dan Mason Telemedicine Support Agent Talk — Notes

Excellent talk by Dan Mason of Stride AI, doing client work for Avila, from AI Engineer World Fair 2025 — thank you to all of them for making this available!

Summary:
- Replaced an existing and highly templatized telemedicine agent support workflow
- Human-in-the-loop flow — AI agents respond in most cases but are able to quickly and efficiently involve a human when needed.
- PM (with engineering background) + 2 engineers built app in a “few months” that would have taken a year. PM used Cline.
- Claims ~10x productivity from prior support workflow!
Observations / insights:
- Rigorous process documentation, with business logic written in English rather than code. He mentioned “1000 lines of english / 1000 lines of code.” — I suspect it’s actually a lot more English, since markdown would consider a whole paragraph 1 line. Documentation was also very well structured, like engineering source code.
💡 Companies with existing processes documented, a la the GitLab Handbook, will have advantages bringing online agents within their company context. More on these practices in TCGen Document Document.
- Smart division of responsibilities between virtual_oa and evaluator — one to do the work and one to be skeptical. Mitigates the overconfidence of the first LLM with the second; both know the existence of the other and criteria used. evaluator confidence scores combine reaction to virtual_oa’s logic and overall complexity of the case.
- An AI (Cline) was able to bring online a new potential protocol (ozempic) with just a new English document, no additional code. Shows tremendous scalability and benefit of separating business logic from code.