Home The signal Anatomy of an agent Reference architecture Risk-Tiers Guardrail stack Governance in action Best practices Standards and crosswalk Implementation From the field Roadmap Companion toolkit Straight answers Glossary References
03 · Reference architecture

The agentic platform, end to end

Seven horizontal planes carry the request. Two vertical rails carry the controls. The rails are the point: governance and evidence are not a plane you pass through once, they touch every plane on every call.

GOVERNANCE RAIL EVIDENCE RAIL L1 · experience & entry L2 · policy & control plane L3 · orchestration L4 · agent runtime L5 · model plane L6 · tool & integration L7 · knowledge & data
Enterprise agentic AI reference architecture the structure is vendor-neutral · the platforms named on each component are worked examples for an Azure and Claude estate, not recommendations
Governance rail
Agent registry
identity, owner, purpose, tierEntra Agent ID + CMDB
Risk tiering
autonomy level per use caseinternal model + GRC tool
Policy as code
versioned, tested, deployedOPA / Cedar in Git
Agent & model cards
purpose, limits, evals, ownerGit + Foundry
Approval workflow
design review to production gateGitHub Environments · Azure DevOps
Change & lifecycle
promote, deprecate, retireServiceNow
Control mapping
ISO 42001 · NIST AI RMF · EU AI ActPurview Compliance Manager
L1Experience & entry planewhere work arrives
Assistant UIchat and copilot surfacesTeams · Copilot Studio · custom web
Business appsembedded agent actionsDynamics 365 · ServiceNow · in-house
IDE & dev toolscoding agentsClaude Code · GitHub Copilot · VS Code
Event triggersqueues, schedules, webhooksEvent Grid · Service Bus · Logic Apps
Agent APIsystem-to-system callsAzure API Management
intent
L2Policy & control planeone authoritative policy system · many enforcement points
Policy administrationauthor, review, version, testGit + OPA bundles · Azure Policy
Policy decisionallow · block · redact · escalateOPA or Cedar · custom PDP service
Identity & entitlementagent identity + delegated principal scopeEntra Agent ID · Entra ID · Workload ID Federation
Deterministic limitsallow-lists, value caps, tenancy, egresspolicy bundle · APIM · network policy
Screening servicesinjection, PII, safety — supplementaryAzure AI Content Safety · Prompt Shields · Presidio
Budget & quotatokens, spend, steps, wall clockAPIM policies · Azure Monitor
Data boundarytenancy, residency, classificationMicrosoft Purview · private endpoints
Kill switchper agent, per tool, globalApp Configuration flags · LaunchDarkly
authorized intent
L3Orchestration planewho does the work, in what order
Planner / routertask decompositionSemantic Kernel · LangGraph · Claude Agent SDK
Supervisor agentdelegation, arbitrationAzure AI Foundry Agent Service
Workflow enginedurable state, retries, compensationDurable Functions · Logic Apps · Temporal
A2A peerscross-team, cross-vendor agentsA2A protocol v1.0
Human-in-the-loopapproval queue, takeoverTeams Adaptive Cards · ServiceNow
Guardian agentmonitors other agents — detection, not boundarycustom on Foundry + Defender signals
plan · step
L4Agent runtime planethe loop, isolated per task
Reasoning loopplan, act, observe, reviseClaude Agent SDK · Semantic Kernel
Context assemblercompaction, pruning, note-takingcustom · Semantic Kernel
Memory serviceworking + durable, scopedCosmos DB · Redis Enterprise · AI Search
Prompt registryversioned, reviewed, testedGit + Foundry prompt assets
Skill libraryreusable, certified patternsprivate MCP servers · internal package feed
Sandbox executorephemeral, no standing credsContainer Apps dynamic sessions · ACI
inference request
L5Model planecapability, routed by cost and risk
Model gatewayrouting, failover, cachingAzure AI Foundry · APIM AI Gateway · LiteLLM
Frontier modelshard reasoning stepsClaude Opus · Claude Sonnet via Foundry or Bedrock
Small / fast modelsclassification, extraction, costClaude Haiku · Phi · distilled models
Tuned modelsdomain-specific behaviourAzure OpenAI fine-tunes · custom adapters
Embeddings & rerankretrieval qualityAzure AI Search vectorizer · Cohere rerank
Safety classifierscontent and risk scoringAzure AI Content Safety · custom
tool call
L6Tool & integration planeevery path to the outside world
MCP server registrysigned, versioned, scannedprivate registry + Azure Container Registry
Tool brokerper-call authorization, token exchangecustom on APIM + Workload ID Federation
API gatewaycontracts, rate limits, mTLSAzure API Management
Event busasync actions, replayService Bus · Event Hubs
Write-back adaptersidempotent, reversibleAzure Functions · Logic Apps connectors
Legacy connectorsmainframe, ESB, fileLogic Apps · Host Integration Server · MuleSoft
grounded data
L7Knowledge & data planewhat the agent is allowed to know
Vector & hybrid searchchunk-level permissionsAzure AI Search
Ontology / knowledge graphentities, relationships, meaningCosmos DB Gremlin · Neo4j · RDF store
Master & reference datathe trusted identifiersProfisee · Informatica MDM · CluedIn
Systems of recordcore platformsSAP · Dynamics 365 · core banking or policy admin
Feature & metric storeconsistent numbersMicrosoft Fabric · Databricks
Catalog & lineageclassification, provenanceMicrosoft Purview · Collibra
Evidence rail
Distributed tracing
OTel GenAI conventionsAzure Monitor · App Insights
Decision log
append-only, tamper-evidentData Explorer · immutable blob
Eval harness
offline, online, adversarialFoundry evaluations · Promptfoo
Cost & token telemetry
per agent, per task, per tenantCost Management + APIM metrics
Drift & quality monitors
output, retrieval, judge driftAzure Monitor + custom
Incident & rollback
replay, contain, remediateAzure DevOps · PagerDuty
Audit & reporting
regulator-ready extractsPurview · Fabric
enforcement point orchestration evidence capability platform example scroll horizontally on narrow screens
RULE 01

One authoritative policy system, no ungoverned paths

Author policy once, version it centrally, and enforce it at every point that touches the outside world. A single physical gateway carrying all model, agent, tool and data traffic becomes a throughput bottleneck, an availability risk and an awkward trust boundary. What must be singular is the policy, not the pipe.

RULE 02

The agent never holds the credential

The tool broker exchanges the agent's identity plus the delegated principal scope for a short-lived, audience-bound token at the moment of the call. Nothing standing, nothing shared, nothing that outlives the task, and no token received from a caller is passed through to a downstream service.

RULE 03

Evidence is written before the action

The intent, the policy decision and the inputs are logged before the tool executes. If logging fails, the action fails. That ordering is what makes an audit trail defensible rather than best-effort.

One authoritative policy system, many enforcement points the four roles a control plane actually plays, drawn separately Recommendation
PAP · POLICY ADMINISTRATION POINT author · review · test · version · publish the bundle SIGNED BUNDLE PDP · POLICY DECISION POINT evaluates identity · action · resource · context returns allow · block · redact · escalate, plus the deciding rule and bundle version PIP · POLICY INFORMATION POINTS identity directory & entitlements agent registry · owner · tier data classification & residency consent & customer preferences risk class & transaction history FACTS EVIDENCE decision written before the action append-only · bundle version attached DECISIONS REQUESTED AND RETURNED Agent gateway entry, budget, identity Azure API Management Tool broker per-call authz, tokens APIM + Workload ID Memory service scope, provenance, TTL Cosmos DB · Redis Retrieval layer chunk-level filtering Azure AI Search Workflow engine approvals, compensation Durable Functions Egress proxy network deny by default Azure Firewall PEP · POLICY ENFORCEMENT POINTS — MANY, BY DESIGN. THE POLICY IS SINGULAR; THE PIPE IS NOT.
authoring decision and enforcement evidence inputs to a decision
Why this beats a single gateway. Enforcement has to happen where the resource lives. Chunk-level retrieval filtering can only be done by the retrieval layer. Memory scoping can only be done by the memory service. Routing all of it through one process to satisfy a diagram adds a hop, a queue and a single point of failure without adding a control.
What makes it coherent. One policy repository, one bundle format, one decision service contract, and conformance tests that prove each enforcement point evaluates the bundle the same way. Every decision records the bundle version, so a run can be replayed against the policy that was live at the time. Decisions may be cached; the staleness window is configured and logged.
One agent request, one lifecycle the same nine checkpoints on every call, whatever the use case
REQUEST DECIDE ACT 1 · Authenticate agent ID + user scope 2 · Screen input injection, PII, intent 3 · Bind budget steps, tokens, spend 4 · Assemble context permission-filtered 5 · Plan model proposes action 6 · Authorize the action allow · block · redact · escalate to a human Human approval required above the tier line Deny and explain reason returned, event recorded 7 · Execute short-lived token, sandbox 8 · Screen output grounding, leakage, safety 9 · Record evidence trace, decision, cost, outcome Feed evaluation regression set, drift, incidents EVIDENCE CLOSES THE LOOP · EVERY RUN IMPROVES THE NEXT RELEASE

Where would you start?

If you are standing up an agent platform, tightening the controls on one you already have, or preparing for an audit that now includes agents, I am happy to look at it with you.