Zynomi Agent Fabric (ZAF)
Zynomi Agent Fabric (ZAF) — the platform a fleet of Zynomi agents runs on, in every department, grounded in a clinical knowledge engine built from a clinical ontology and GraphRAG, with every agent output carrying provenance and an accept/reject disposition.
How It's Designed
Four checkable architecture facts
ZAF is described by its architecture, not by promises. These are the design facts every part of the platform is built around.
Agent + human-approval loop
Every workflow pairs an agent with a human approval step. Agents do the data work; your team reviews and approves every result before it counts.
Ontology + GraphRAG grounding
Agents are grounded in a clinical knowledge engine built from a designed clinical ontology and GraphRAG — they retrieve from governed structures, not free-form guesswork.
Semantic layer + MCP interface
A governed semantic layer defines every metric once; MCP is the cross-system interface that lets agents work across the systems trials already run on.
Provenance + disposition on every output
Every agent output carries provenance — where it came from and how it was derived — and an accept/reject disposition recorded with the result.
The Fleet
A crew of bounded agents, not one chatbot
ZAF is designed as a supervisor that owns no data tools, routing questions to bounded specialists — each seeing exactly one engine. Tools compute, the model narrates: agents never re-derive a number and never invent an edge.
Supervisor
Owns no data tools — classifies the question and routes it to bounded specialists
Trial Metrics Analyst
Governed KPIs only
Semantic Layer
Centrally defined metrics
Study Historian
Fixed-shape Cypher via MCP
Temporal Graph
Relationships + time
Operations Agent
Read-only lookups
Transactional Store
Current operational state
Data Steward
Quality & lineage checks
Quality / Lineage
Tests and lineage metadata
One engine per specialist
Each agent sees only its own context's tools
Read-only by construction
One gate, tested in CI — not promised in a prompt
Tools compute, the model narrates
Agents never re-derive a number, never invent an edge
The Knowledge Engine
The third data layer
Trials already have transactional and analytical layers. ZAF's clinical knowledge engine adds the third: a temporal knowledge graph that answers the questions the other two cannot.
Transactional
"What is happening now?"
The operational databases the trial runs on — visits, forms, consents, queries — optimized for the current state of every record.
- Current subject status
- Open queries and tasks
- Live operational lookups
Analytical
"What happened at scale?"
The warehouse and governed semantic layer — aggregates, trends, and KPIs computed from centrally defined metrics.
- Enrollment rates and trends
- Governed KPIs
- Cross-study aggregates
Temporal Knowledge Graph
"What is connected, how did it evolve, and why?"
The third layer: entities and relationships with time on every edge — the connective tissue the first two layers cannot express.
- Amendment provenance
- Affected sites and activities
- Before/after history of any change
“Explain everything around Amendment 4 — who requested it, who approved it, which sites were affected, what happened after.”
Designed to be answered in one graph traversal — instead of days of joining audit tables.
Reference Architecture
Two paths, one interface
Governed numbers travel the analytical path; relationships and time travel the graph path. Both converge at the MCP layer that serves the agents.
Clinical Applications
The systems the trial already runs on
Analytical Path — Governed Numbers
Transactional DB
Operational records
Warehouse
Modeled, historized data
Semantic Layer
Centrally governed KPIs
Graph Path — Relationships + Time
Business Events
Stream / CDC
Graphiti
Extraction, resolution, temporal processing
FalkorDB
Temporal knowledge graph
MCP Layer
Both paths served through one protocol
AI Assistant
Agents run on Amazon Bedrock (AWS)
Where We Are
Built in the open, described honestly
ZAF is being built ground-up in 2026. Nothing on this page is a shipped-product claim — it describes the architecture the platform is designed as, and the plumbing that already exists underneath it:
We publish design facts you can check — not performance numbers we haven't measured. Compliance capabilities are designed for and designed to support the relevant regulations.
The Design Behind ZAF
Read the Zynomi white paper
Living Clinical Trial Knowledge Graph — a temporal, relationship-aware semantic architecture for clinical trial operations. The vendor-neutral design behind ZAF's clinical knowledge engine and agent fleet.
Living Clinical Trial Knowledge Graph
A temporal, relationship-aware semantic architecture for clinical trial operations — a Zynomi white paper (PDF).