AI agents
Narrow mandate. Grounded context. Structured output. A human at the end of every one.
Overview
Agents with a narrow mandate and a human at the end
An AI agent in healthcare is only safe when its mandate is small, its context is retrieved rather than remembered, its output is structured, and a human closes the loop.
Design rules
- One narrow mandate per agent
- Grounded, retrieved context
- Structured, validated outputs
- Deterministic fallback
- Global kill-switch
Supervision
- Every output routed to a human queue
- Explainable rationale attached
- Reversible by design
- Immutable audit log
- No silent changes
Applications
- Validation & data quality
- Triage and prioritisation support
- Referral follow-up
- Reporting drafts
- Research and insight assistance
The boundary
What an agent never does
No agent prescribes, diagnoses, or closes a clinical loop on its own. The validation agent flags and a human corrects. The triage agent suggests an order and a clinician confirms. The reporting agent drafts and an officer signs. That boundary is enforced in the orchestrator, not left to good intentions.
Scope an agent.
Start with the task, the failure mode, and who reviews the output.