The Agentic Stack
An operating system for agents, built on the "Trust Triangle".
Layered Safety Architecture
Three distinct layers ensuring interoperability, intelligence, and safety.
Layer 3: The Guardian Layer (The "Brakes")
Input
Guardrails (PII)
Output
Guardrails (Clinical)
Hallucination Check
Critic Agent
Layer 2: The Orchestration Engine (The "Brain")
LangGraph Orchestrator
Vector Memory (RAG)
State Management
Reasoning Engine
Layer 1: The Data Foundation (Interoperability)
FHIR Connectors
SMART on FHIR
EHR Adapters (Epic/Cerner)
HAPI FHIR Server
Component Deep Dive
Technical specifications for the core system layers.
Layer 1: Data Foundation
Agents need a standardized way to "read" and "write" to hospital systems.
- Standard: FHIR (Fast Healthcare Interoperability Resources). All inputs/outputs map to FHIR resources.
- Integration: Connectors for major EHRs (Epic, Cerner) using SMART on FHIR.
- Vendor Agnostic: Agents talk to a generic "Connector", not specific APIs.
Layer 2: Orchestration Engine
The "Brain" that manages agents and decides which agent performs which task.
- Framework: Uses LangGraph or Temporal.io for long-running workflows (e.g., referrals).
- Memory Store: Vector database (Qdrant, Weaviate) for "Long-Term Memory" of patient history.
- State Management: Ensures the system never "forgets" state over days or weeks.
Layer 3: Guardian Layer
The "Brakes" that enforce safety and compliance.
- Input Guardrails: Detects PII and "jailbreak" attempts before LLM processing.
- Output Guardrails: Verifies medical accuracy against clinical guidelines.
- Hallucination Check: A secondary "Critic Agent" reviews actions before execution.
Built on Open Standards
FHIR
Data Standard
Data Standard
LangGraph
Orchestration
Orchestration
Qdrant/Weaviate
Vector Memory
Vector Memory
Docker
Containerization
Containerization
Guardian
Safety Layer
Safety Layer
Python 3.11+
Core Runtime
Core Runtime