The Agent Layer Is the Hard Part: Why Most Agentic AI Pilots Don't Reach Production

dehakuran.com · May 2026 · 3 min read

A swarm of agents, you will have. Handle them, how will you? Without a solid Agentic OS underneath, hmmmm?


Agents, I see. Agents everywhere, I see. Yet bothered by the coming problem, no one seems to be.

Stop asking which agent to buy or build. Start asking what your agent layer looks like.

The model is the easy part. The layer underneath is the hard part — the one that lets agents remember, coordinate, call tools, and not lose the plot at step seven.

Call it what you want. Most are landing on Agentic OS.

Where It's Getting Tested Hardest: Healthcare

A few signals from just the last few months:

  • Hackensack Meridian runs an agent named Erin for post-discharge follow-ups — symptoms, meds, appointments, escalation. Live, in production.
  • Microsoft Healthcare Agent Orchestrator is being piloted at Stanford for tumor boards. Specialized agents for pathology, radiology, staging, trial-matching. A coordinator pulls it together for the oncologist.
  • The UK launched TrustX Health in December — a national framework to verify and safely deploy agentic AI across the NHS.

But Reality Kicks In With Data

  • 43% of health systems are piloting agentic AI.
  • Only ~3–4% have it in live workflows.
"That gap isn't a model problem. It's an OS problem."

What an Agent Layer Actually Has to Do

  • FHIR interoperability with legacy EHRs.
  • Identity and governance for agents, not users.
  • HIPAA-aware architectures that keep PHI out of the inference plane.
  • Orchestration that survives a real hospital, not a demo.

The Real Question

So as I said: stop asking which agent to buy. Start asking what your agent layer looks like.

The orgs that figure out the OS layer first won't have better agents. They'll have agents that actually ship.

The model is commoditizing. The plumbing is where the edge lives.


May the source be with you. Not sure how much will be left of it from agents though. 😅

Source: A Microsoft / Health Management Academy survey, published in NEJM AI. Views are my own and do not represent any commercial entity.

Frequently Asked Questions

What is an Agentic OS, and why is it the bottleneck for agentic AI?

An Agentic OS is the layer underneath the model that lets agents remember state, coordinate with other agents, call tools, and not lose the plot at step seven. It's the bottleneck because the model is commoditizing — frontier models converge within months — while the OS underneath has to handle agent identity and governance, tool orchestration, and architectures that survive a real production environment. The model is the easy part; the layer underneath is the hard part.

Why are so few health systems running agentic AI in production?

A Microsoft / Health Management Academy survey published in NEJM AI found 43% of health systems are piloting agentic AI but only ~3–4% have it in live workflows. The gap is not a model problem — it's an Agentic OS problem: FHIR interoperability with legacy EHRs, identity and governance for agents (not users), HIPAA-aware architectures that keep PHI out of the inference plane, and orchestration that survives a real hospital, not a demo.

Which healthcare organizations have agentic AI in production today?

Hackensack Meridian runs an agent named Erin for post-discharge follow-ups — symptoms, medications, appointments, and escalation — live in production. Microsoft Healthcare Agent Orchestrator is being piloted at Stanford for tumor boards, with specialized agents for pathology, radiology, staging, and trial-matching coordinated for the oncologist. The UK launched TrustX Health in December 2025, a national framework to verify and safely deploy agentic AI across the NHS.

AI AgentsHealthcareEnterprise

Deha Kuran

AI Executive, Engineer, and Evangelist. Head of AI Business Operations at Philips.

Follow the thinking on LinkedIn →