Architecture
How it comes together, end to end.
One reference implementation, from the data to the agent to the security operations center, on the cloud you already run.
The full picture
Three disciplines, one system.
Most programs fund AI, data and security separately. They only work when designed together, and held together by orchestration and an operating model: the process, control plane, release gates and accountable owners that run them.
One system, three views
This diagram is the main view. The other two describe the same system for a different purpose. None is a separate framework to buy or implement.
How the three views fit together. The three disciplines set what must be true: Trusted Data, Tested Capability and Bounded Access. The Four-Part System assigns the architectural job behind each one. Trusted Data is the brain’s work, and Bounded Access is the immune system’s. Tested Capability is what the four parts serve: agents and models reach work through the nervous system, run in the body, and are proven at release gates recorded in the control plane. The four phases, Frame, Prove, Build, then Operate and transfer, are how an engagement gets one workflow there.
The architecture
The agentic enterprise, layer by layer.
People and agents work in redesigned processes on top of five layers. Security frames every layer, and orchestration and operation run beside them. The names match the capability map and the product layers below.
Security, across every layer · identity for people and agents, labels, enforcement at the data, monitoring
People and agents
Redesigned business processes. Agents act within the authority of the person they serve, and a person stays in the loop for consequential actions.
AI and experience
ReasoningToolsMemoryOrchestrationGuardrailsEvaluationCatalogObservability
Foundation modelsAdapted modelsReportingSelf-service
Access and interfaces
Data APIsGoverned measuresAI gatewayAgent toolsRetrievalCommand path for writes
Meaning and knowledge
Ontology and knowledge graphSemantic layerMetadataMaster dataSearch
The ontology says what is meant; the semantic layer computes the number. Context for the answer, never a grant of access.
Data platform
Data productsDomain platformsPipelinesStorageVector storage
Access is enforced here: masking and row access, every time.
Sources
Operational applicationsDocumentsExternal data
Orchestration and operation
- Release gates and evaluation
- DataOps, DevOps, model operations
- Quality and lineage
- FinOps: cost per completed task
- Request evidence: who asked, which definition, measure and data versions, which policy decision, what happened
Operating model, underneath
Accountable owners, decision rights, skills and funding for running, not only building.
Vendor neutral. Four requirements anchor this design: identity for every agent, meaning separate from access authority, governed numbers from fixed measures rather than the model’s own arithmetic, and evidence that lets an authorized reviewer reconstruct a request.
On your platform
The same architecture on AWS, Azure, Google Cloud, Snowflake, Databricks or a mixed estate.
24 layers across the four parts and the control plane, 161 named services, each with its release status checked on 13 November 2026. Pick a platform, or the mixed estate most enterprises actually run, then click a layer, step through a flow, or compare platforms side by side.
| Layer | AWS | Microsoft Azure | Google Cloud | Snowflake | Databricks | Mixed estate (example) |
|---|---|---|---|---|---|---|
| Catalog, metadata and lineage | AWS Glue Data Catalog GA Amazon SageMaker Catalog GA | Microsoft Purview Unified Catalog GA OneLake catalog GA | Knowledge Catalog (formerly Dataplex Universal Catalog) GA | Snowflake Horizon Catalog GA | Unity Catalog GA | data.world (ServiceNow) GA |
| Semantic layer and metrics | No native universal semantic layer Gap | Power BI semantic models GA Fabric IQ (workload) GA | Looker semantic layer (LookML) GA BigQuery measures Preview | Semantic views GA | Unity Catalog metric views GA | Semantic views GA dbt Semantic Layer GA |
| Ontology and knowledge graph | Amazon Neptune Database GA Amazon Neptune Analytics GA | Fabric IQ ontology Preview Graph in Fabric GA | Spanner Graph GA BigQuery Graph Preview | No native graph store Gap | Genie Ontology Preview | data.world knowledge graph GA |
| Knowledge retrieval and vector search | Amazon Bedrock Knowledge Bases GA Knowledge Bases GraphRAG with Neptune Analytics GA Amazon S3 Vectors GA | Azure AI Search GA Foundry IQ GA | Vertex AI Search (Gemini Enterprise Agent Platform) GA BigQuery vector search GA | Cortex Search GA | Mosaic AI Vector Search GA Knowledge Assistant GA | Cortex Search GA Amazon Bedrock Knowledge Bases GA |
| Master data and entity resolution | AWS Entity Resolution GA | No native MDM service Gap | No native MDM service Gap | No native MDM service Gap | No native MDM service Gap | Partner MDM Gap |
| SaaS and enterprise source connectors | AWS Glue zero-ETL for applications GA Amazon AppFlow GA | Fabric Data Factory connectors and Copy job GA Fabric mirroring (SAP Datasphere, Oracle, Dataverse) GA | Integration Connectors GA Cortex Framework for SAP and Salesforce GA | Snowflake Openflow GA | Lakeflow Connect GA | Snowflake Openflow GA |
| Change data capture and streaming | AWS Database Migration Service (CDC) GA Amazon Kinesis Data Streams and Amazon MSK GA Zero-ETL from Aurora, RDS and DynamoDB GA | Fabric mirroring for databases GA Azure Event Hubs GA | Datastream GA Pub/Sub and Managed Service for Apache Kafka GA | Snowpipe Streaming GA Openflow database connectors Confirm | Lakeflow Connect database connectors GA Structured Streaming GA | AWS Database Migration Service (CDC) GA Apache Kafka GA |
| APIs, events and workflow | Amazon API Gateway GA Amazon EventBridge and AWS Step Functions GA | Azure API Management GA Azure Logic Apps and Event Grid GA | Apigee GA Workflows and Eventarc GA Reverse ETL for BigQuery Preview | Snowflake SQL API and external access integrations GA | Databricks Apps and Lakeflow Jobs GA Lakebase GA | Hasura DDN GA Amazon API Gateway GA |
| Agent tools, gateway and MCP | Amazon Bedrock AgentCore Gateway GA AWS MCP Server GA | MCP governance in Azure API Management GA Azure MCP Server GA Azure Logic Apps MCP Server GA | Google-managed MCP servers GA Agent Gateway (Gemini Enterprise Agent Platform) Confirm | Snowflake-managed MCP server GA MCP connectors for Cortex Agents Preview | Unity AI Gateway MCP services GA Databricks managed MCP servers Preview | Amazon Bedrock AgentCore Gateway GA |
| Storage, open tables and history | Amazon S3 and S3 Tables (Apache Iceberg) GA | OneLake and Azure Data Lake Storage Gen2 GA | Cloud Storage and BigLake Iceberg tables GA | Apache Iceberg tables GA | Delta Lake and Iceberg managed tables GA | Apache Iceberg tables GA |
| Warehouse, lakehouse and operational stores | Amazon Redshift, Athena and EMR GA | Fabric Data Warehouse and Lakehouse GA | BigQuery, Spanner and AlloyDB GA | Snowflake virtual warehouses GA | Databricks SQL GA | Snowflake virtual warehouses GA |
| Pipelines and transformation | AWS Glue ETL and Amazon MWAA GA | Fabric pipelines, Dataflow Gen2 and materialized lake views GA | Dataform, Dataflow and Cloud Composer GA | Dynamic Tables GA | Lakeflow Declarative Pipelines GA | dbt GA Apache Airflow GA |
| Data products and sharing | SageMaker Catalog data products GA Redshift data sharing and AWS Data Exchange GA | Purview Unified Catalog data products GA | Knowledge Catalog data products GA BigQuery sharing (Analytics Hub) GA | Secure Data Sharing and Snowflake Marketplace GA | OpenSharing (Delta Sharing) and Clean Rooms GA | Secure Data Sharing and Snowflake Marketplace GA |
| Models and AI platform | Amazon Bedrock GA | Microsoft Foundry Models GA | Model Garden (Gemini Enterprise Agent Platform) GA | Cortex AI and AISQL GA | Model Serving and Foundation Model APIs GA | Amazon Bedrock GA Cortex AI and AISQL GA |
| Agent runtime and memory | AgentCore Runtime and Memory GA Amazon Bedrock Agents GA | Foundry Agent Service GA | Agent Runtime, Sessions and Memory Bank GA | Cortex Agents and Snowflake Intelligence GA | Agent Bricks and Genie Agents Confirm Managed agent memory and sessions Beta | AgentCore Runtime and Memory GA |
| Observability, evaluation and cost | AgentCore Observability and Evaluations GA Amazon CloudWatch GA | Foundry Control Plane observability and evaluations GA Azure Monitor GA | Cloud Trace, Monitoring and agent evaluation Confirm | Account usage views and AI observability Confirm | MLflow tracing and system tables GA | OpenTelemetry GA AgentCore Observability and Evaluations GA |
| Identity for people and agents | AWS IAM Identity Center and AgentCore Identity GA | Microsoft Entra ID and Entra Agent ID GA | Cloud IAM GA Agent Identity Confirm | Role-based access control and OAuth GA | Unity Catalog identities and service principals GA | Okta GA Microsoft Entra ID and Entra Agent ID GA |
| Access policy at the data | AWS Lake Formation and Amazon Verified Permissions GA | OneLake security GA | BigQuery row and column security, policy tags and VPC Service Controls GA | Masking and row access policies GA | Attribute-based access control (ABAC) GA | Masking and row access policies GA |
| Classification, DLP and DSPM | Amazon Macie GA | Microsoft Purview Information Protection, DLP and DSPM GA | Sensitive Data Protection GA | Sensitive data classification GA | Data Classification GA | Cyera GA Microsoft Purview Information Protection, DLP and DSPM GA |
| AI guardrails and action policy | Amazon Bedrock Guardrails GA Policy in AgentCore GA | Azure AI Content Safety and Prompt Shields GA Content safety for MCP and A2A in API Management GA | Model Armor GA | Cortex Guard GA | Unity AI Gateway guardrails Confirm | Amazon Bedrock Guardrails GA Policy in AgentCore GA |
| Security operations (SIEM and SOAR) | AWS Security Hub, Amazon GuardDuty and Amazon Security Lake GA | Microsoft Sentinel and Defender for Cloud GA | Google Security Operations and Security Command Center GA | Trust Center GA | Lakewatch Confirm | Your SIEM and SOAR Confirm |
| Keys and confidential computing | AWS KMS and AWS Nitro Enclaves GA | Azure Key Vault and Azure confidential computing GA | Cloud KMS and Confidential Computing GA | Tri-Secret Secure GA | Customer-managed keys GA | AWS KMS and AWS Nitro Enclaves GA Tri-Secret Secure GA |
| Agent registry and release control | Amazon Agent Registry GA | Microsoft Agent 365 and Foundry Control Plane inventory GA | Agent Registry (Gemini Enterprise Agent Platform) Confirm | Cortex Agent versioning GA | Unity AI Gateway registry GA | Amazon Agent Registry GA ServiceNow change and incident management GA |
| Audit and governance evidence | AWS CloudTrail and AWS Audit Manager GA | Microsoft Purview Audit and Compliance Manager GA | Cloud Audit Logs and Assured Workloads GA | Access History GA | Audit log system tables GA | AWS CloudTrail and AWS Audit Manager GA Access History GA |
Examples, not endorsements. GA means generally available; Preview and Beta are not production commitments; Confirm means vendor sources disagree or no dated source was found; Gap means there is no native service. Status reflects public vendor sources on the date shown. Download the catalog (JSON). One page per platform: AWS, Azure, Google Cloud, Snowflake, Databricks. Richard has worked closely with several vendors named here as a customer and practitioner. No vendor reviewed or paid for this content.
Use case · reference example
A claims assistant, traced end to end.
A claims examiner asks who is connected to an assigned case, its status, and which procedure applies. This read-only architecture pattern supports the claims preparation guide: team A, group disability, region 2. Events within the assigned case are included; prior-claims history, internal notes, decisions, payments and customer contact are excluded from its narrowed release.
- 01
The examiner signs in through single sign-on. The assistant runs under its own agent identity, registered to the claims owner, with a defined role.
- 02
The request carries a short-lived token scoped to this examiner and this agent: only what both are allowed.
- 03
A fixed sequence resolves claim status, the related parties and the applicable procedure. Cross-case search and note tools are unavailable in the narrowed runtime.
- 04
Claim status comes from a certified data product through a narrow API. Tag-based masking hides fields this examiner may not see.
- 05
Procedure documents were classified and labeled by the DSPM tool. Retrieval returns only documents this examiner may read, for this tenant.
- 06
One document contains hidden instructions to email the file outside the company. The agent has no tool that can send externally; the attempt is refused and flagged.
- 07
The answer separates facts from interpretation, cites its sources and the definition version it used, and flags an unresolved party match for the examiner.
- 08
Every step is traced with the agent’s identity, the user, the tools, the data labels and each policy decision, and streamed to the SOC.
- 09
Days later, the agent starts requesting claims outside the examiner’s assigned book. Each call is denied; the repeated denials alert the SOC, and a playbook suspends the agent identity and opens an incident for the owner.
- 10
When an examiner leaves, the identity provider removes access, and the revocation test shows the assistant stops using it within the tested window.
- 11
The trace, the refused instruction, the incident and the revocation test become release evidence. The owner decides whether to restore the agent, and on what conditions.
The steps illustrate architecture and incident handling, not a client deployment or an executed test. The versioned workflow record and completed release decision define the worked guide’s accepted scope and held dependencies.
Orchestration
The process and technology that hold it together.
Control plane
One record of what was approved: versions, policies, owners and the conditions of each release.
Release pipeline
Every change, to data, model, prompt, tool or policy, passes the same gates.
Policy propagation
Each rule declares its source, targets, latency, drift check, rollback and owner.
Observability
One request ID from the user through every agent, tool and source, into the SOC.
Runbooks and recovery
Tested restriction, rollback and revocation, run by your own team.
Operating model
The data group owns meaning, the cyber group owns labels and enforcement rules, the AI owner owns the capability, and the business owns the outcome.
Go deeper
The detail behind the architecture.
Open what you need: the capability map for assigning owners, the costs that arrive after approval, and how agentic delivery stays safe.
The capability map: 63 capabilities, mapped to the same layers
The ownership view of the same system shown at the top of this page.
Ownership gaps stall more programs than technology does. The worksheet lists every capability with a one-line description and blank columns for current owner, accountable owner, decision rights and gaps. It assigns no owners: the right owner depends on your organization.
AI and experience
The capability the four parts serve · 13
Reporting and analytics, Self-service exploration, Data-driven applications, Data science sandbox, Prompt and instruction library, Foundation models, Model adaptation, Agents, Agent orchestration, Agent memory, Agent evaluation, Agent catalog, Agent observability
Access and interfaces
The nervous system · 6
Data APIs, Virtualization and federation, AI gateway, Retrieval services, Agent tool interfaces, Command path for writes
Meaning and knowledge
The brain · 8
Metadata, Ontology and knowledge graph, Semantic layer, Information architecture, Master and reference data, Chunking and embeddings, Data marketplace and search, AI asset search
Data platform
The brain’s data products · 12
Batch integration, Streaming and messaging, Pipeline orchestration, Data pipelines, Schema detection, Data modeling, Data preparation, Data storage, Vector storage, Cloud hosting, Containers, Accelerated compute
Sources
Systems of record · 3
Purpose-built databases, Operational applications, External and web data
Security, across every layer
The immune system · 7
Identity and access for people, Identity for agents, Single sign-on and zero trust, Classification and labels, Data protection, Monitoring and logging, Threat detection and response
Orchestration and operation, across every layer
The body and control plane · 10
Data quality, Lineage, Data architecture standards, Responsible AI, DataOps, DevOps, Model and AI operations, FinOps, Release gates, Request evidence
Operating model
Underneath all four · 4
Data and AI strategy, Upskilling and training, Research and talent programs, External workforce
Total cost of ownership: twelve recurring costs after approval
Build, licenses, vendor fees, integration and pilot compute are visible at approval. Twelve recurring costs arrive after it, and the architecture above is what keeps them in bounds.
Agentic delivery: faster code still needs authority, independent checks and learning
Coding agents can draft, test and execute changes within bounded tools and identities. The loop starts from governed intent, validates independently of the agent that made the change, releases under named authority and feeds what operation reveals back into intent.
The paper
The full reference implementation.
The paper sets out the architecture diagrams and product detail layer by layer, the security model for agents and the security operations center, and two use cases traced end to end: anti-money-laundering alert investigation at a bank, and this claims workflow at an insurer. Twenty-three pages, free, no form.
Next step
See how this would fit your stack.
Set up time with Richard to discuss one workflow, what is blocking it and whether there is a useful next step. The first conversation is free.