I have been looking at healthcare through the systems I build with teams, and I keep reaching the same conclusion: models are ready for useful work, but the infrastructure around them is not.
My definition of agentic healthcare is practical. Ground the workflow in evidence, keep accountable people in the loop, and let machines handle the repetitive work that never needed a clinician’s attention.
Research before reasoning
When I think about a research agent, I start with the records and guidelines a care team already has. The problem is usually not a lack of answers. It is that the right context is scattered, slow to retrieve, or impossible to inspect in one place. A useful agent gathers that context, cites it, and exposes what it cannot verify.
For high-stakes work, my rule is simple: look it up, do not make it up. Retrieval comes before synthesis, and citations stay attached to the answer. I cover the same guardrail in Preventing Hallucinations: Building More Honest AI Agents.
CLI-native infrastructure
In the healthcare workflows I design, agents work best with simple, scriptable, auditable tools. I prefer a CLI because it is deterministic, composable, and easy to sandbox.
I use MCP servers only when they add a clear boundary, and I keep them read-only. The agent gathers data; it does not write back to a patient record by default.
The CLI is the core interface I would put underneath healthcare automation. I treat models as pipeline components with strict inputs and outputs, a pattern I cover in Agent-Friendly CLI Tools: From Flaky Agents to Reliable Automation.
Safety is workflow design
My experience building agent workflows has made the distinction clear: safety is system design. A model does not deliver safe healthcare by itself. The workflow does.
I want traceable steps, schema-locked outputs, and rerunnable evaluations before I trust an agent with healthcare work. I unpack that engineering side in Reliable Agents: Engineering Reliable AI Agents.
I treat automatic PII detection and masking as a workflow requirement. Healthcare data can include names, addresses, and SSNs, so an agentic pipeline should strip identifiers before data reaches a model and reconstruct the context only after processing.
For token efficiency, I would remap PII to single-token integers.
I reconstruct the protected context only after model processing.
Schema-locked outputs keep the model from returning arbitrary fields.
I recommend the Anthropic Healthcare and Life Sciences Livestream as a useful public reference for this direction.
The consumer-facing evolution
The ChatGPT Health launch on January 7 gave me another public event to compare with the systems I am building. Users can connect health data from Apple Health, Function, and MyFitnessPal, allowing responses to be grounded in their information. OpenAI says more than 230 million people ask health questions on ChatGPT each week.
The design lines up with principles I use in healthcare work: isolate health conversations, evaluate with clinical input, and keep the agent from diagnosing or treating. The product is positioned as a way to understand patterns and prepare for appointments.
The outcome I want is a patient who arrives with synthesized lab results, relevant research, and clear questions. That makes the patient a better partner in care and gives the clinical workflow better context.
The ChatGPT Health announcement is the public event I used as a reference point for this section.
The future I want
No autonomous medicine. Cooperative systems. Every clinician needs a research agent that keeps records current, finds relevant evidence, and flags what needs human judgment.
This is not a tech problem. It is integration, policy, trust. That is where the work is.
Disclaimer: I am building an auto-coding healthcare platform for a client. That influences my view on where agentic healthcare is heading.