[ { "id": "01-system-context", "title": "AI Agent Orchestration Platform — System Context", "layout": "context", "colWidth": 250, "system": { "label": "AI Agent Orchestration Platform", "sub": "Azure · control + execution" }, "groups": [ { "side": "top", "title": "Identity and governance", "nodes": [ { "id": "entra", "label": "Microsoft Entra ID", "kind": "external", "sub": "SSO · OIDC · agent ID", "rel": "tokens", "dir": "in", "icon": "azure_aad" }, { "id": "purview", "label": "Microsoft Purview", "kind": "external", "sub": "classification", "rel": "labels", "dir": "in", "icon": "azure_AzurePurviewAccount (Analytics)" }, { "id": "sentinel", "label": "Microsoft Sentinel", "kind": "external", "sub": "SIEM", "rel": "security events", "kind2": "async", "icon": "azure_sentinel" } ] }, { "side": "left", "title": "People", "nodes": [ { "id": "dev", "label": "Agent Developer", "kind": "actor", "sub": "SDK · CLI · builder", "rel": "defines agents", "dir": "in" }, { "id": "user", "label": "Business User", "kind": "actor", "sub": "chat · app · API", "rel": "runs agents", "dir": "in" }, { "id": "approver", "label": "Approver", "kind": "actor", "sub": "Teams card", "rel": "approves actions", "dir": "in" }, { "id": "admin", "label": "Platform Admin", "kind": "actor", "sub": "console", "rel": "sets policy", "dir": "in" }, { "id": "sre", "label": "SRE", "kind": "actor", "sub": "on call", "rel": "operates", "dir": "in" } ] }, { "side": "right", "title": "Model providers", "nodes": [ { "id": "foundry", "label": "Azure AI Foundry", "kind": "external", "sub": "GPT · embeddings", "rel": "inference", "icon": "azure_ai-foundry" }, { "id": "fmodels", "label": "Foundry Models", "kind": "external", "sub": "partner catalogue", "rel": "specialist routes", "icon": "azure_foundry-models" }, { "id": "kaito", "label": "Open-Weight Models", "kind": "external", "sub": "AKS · KAITO", "rel": "sovereign routes", "icon": "azure_kubernetes-services" }, { "id": "extllm", "label": "External LLM APIs", "kind": "external", "sub": "egress controlled", "rel": "fallback only", "icon": "fa5_cloud" } ] }, { "side": "bottom", "title": "Tools and systems of record", "nodes": [ { "id": "mcp", "label": "MCP Tool Servers", "kind": "external", "sub": "internal + partner", "rel": "tool calls", "icon": "fa5_plug" }, { "id": "saas", "label": "SaaS Applications", "kind": "external", "sub": "SAP · SFDC · SNOW", "rel": "actions", "icon": "fa5_building" }, { "id": "entapi", "label": "Enterprise APIs", "kind": "external", "sub": "via APIM", "rel": "read and write", "icon": "fa5_plug" }, { "id": "content", "label": "Content Estate", "kind": "external", "sub": "SharePoint · ADLS", "rel": "grounding", "icon": "fa5_folder_open" }, { "id": "teams", "label": "Microsoft Teams", "kind": "external", "sub": "Graph", "rel": "notifications", "dir": "out", "kind2": "async", "icon": "teams" } ] } ], "note": "Model and tool endpoints are reached only through the platform gateway. No agent holds a provider key.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "02-high-level-architecture", "title": "AI Agent Orchestration Platform — High-Level Architecture", "layout": "flow", "chain": true, "canvas": { "width": 1780 }, "stages": [ { "title": "Entry", "nodes": [ { "id": "h-fd", "label": "Front Door", "kind": "integration", "sub": "WAF · anycast", "icon": "azure_AzureFrontDoor (Networking)" }, { "id": "h-apim", "label": "Platform API", "kind": "integration", "sub": "APIM · OIDC", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "h-sdk", "label": "SDK, CLI, Console", "kind": "app", "sub": "Python · .NET · TS", "icon": "python" } ] }, { "title": "Admission", "nodes": [ { "id": "h-adm", "label": "Admission Service", "kind": "app", "sub": "resolve and pin" }, { "id": "h-pol", "label": "Policy Engine", "kind": "security", "sub": "OPA · signed bundle", "icon": "opa" }, { "id": "h-bud", "label": "Budget Guard", "kind": "security", "sub": "cost and token cap", "icon": "azure_cost-budgets" } ] }, { "title": "Orchestration", "nodes": [ { "id": "h-orch", "label": "Durable Orchestrator", "kind": "app", "sub": "Durable Functions", "icon": "azure_functions" }, { "id": "h-sched", "label": "Scheduler", "kind": "app", "sub": "priority classes" }, { "id": "h-coord", "label": "Agent Coordinator", "kind": "app", "sub": "multi-agent patterns" }, { "id": "h-hitl", "label": "Approval Service", "kind": "app", "sub": "external events", "icon": "fa5_user_check" } ] }, { "title": "Agent runtime", "nodes": [ { "id": "h-exec", "label": "Agent Executor", "kind": "app", "sub": "Container Apps", "icon": "agent" }, { "id": "h-ctx", "label": "Context Manager", "kind": "app", "sub": "assemble · compress" }, { "id": "h-grd", "label": "Guardrail Chain", "kind": "security", "sub": "Content Safety", "icon": "azure_content-safety" }, { "id": "h-sbx", "label": "Session Sandbox", "kind": "platform", "sub": "dynamic sessions", "icon": "azure_azure-enclaves" } ] }, { "title": "Model and tool access", "nodes": [ { "id": "h-gw", "label": "AI Gateway", "kind": "integration", "sub": "APIM GenAI policy" }, { "id": "h-tool", "label": "Tool Broker", "kind": "integration", "sub": "MCP · REST" }, { "id": "h-mem", "label": "Memory Service", "kind": "app", "sub": "scoped retrieval", "icon": "fa5_memory" } ] }, { "title": "Evidence", "nodes": [ { "id": "h-state", "label": "Execution Store", "kind": "store", "sub": "Cosmos DB", "icon": "azure_azure-cosmos-db" }, { "id": "h-trace", "label": "Trace and Cost", "kind": "platform", "sub": "Monitor · ADX", "icon": "azure_monitor" }, { "id": "h-audit", "label": "Audit Ledger", "kind": "store", "sub": "immutable blob", "icon": "azure_blobstorage" } ] } ], "edges": [ { "from": "h-orch", "to": "h-state", "label": "checkpoint", "route": "gutter" }, { "from": "h-exec", "to": "h-trace", "label": "spans", "kind": "async", "route": "gutter" }, { "from": "h-hitl", "to": "h-audit", "label": "decision", "kind": "async", "route": "gutter" } ], "note": "One picture for the reader who sees only one. Tenancy, regions and failure paths are later views.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "03-control-and-execution-planes", "title": "Control Plane and Execution Plane — The Governing Boundary", "layout": "nested", "canvas": { "width": 1720 }, "boxes": [ { "title": "Control Plane · what a run is allowed to be", "kind": "cloud", "icon": "azure", "dir": "row", "children": [ { "title": "Registries", "kind": "plain", "nodes": [ { "id": "p-areg", "label": "Agent Registry", "kind": "app", "sub": "versioned specs" }, { "id": "p-wreg", "label": "Workflow Registry", "kind": "app", "sub": "published graphs" }, { "id": "p-treg", "label": "Tool Registry", "kind": "app", "sub": "API Center", "icon": "azure_api-center" }, { "id": "p-mreg", "label": "Model Registry", "kind": "app", "sub": "routes and quotas", "icon": "azure_foundry-control-plane" } ] }, { "title": "Governance", "kind": "plain", "nodes": [ { "id": "p-pol", "label": "Policy Authoring", "kind": "security", "sub": "Rego · versioned", "icon": "opa" }, { "id": "p-idn", "label": "Identity and RBAC", "kind": "security", "sub": "Entra · scopes", "icon": "azure_aad" }, { "id": "p-sec", "label": "Credential Broker", "kind": "security", "sub": "Key Vault refs", "icon": "azure_AzureKeyVault (Security)" }, { "id": "p-cfg", "label": "Config Promotion", "kind": "app", "sub": "App Configuration", "icon": "azure_app-configuration" } ] }, { "title": "Control data", "kind": "plain", "nodes": [ { "id": "p-sql", "label": "Definition Store", "kind": "store", "sub": "Azure SQL · HA", "icon": "azure_azure-sql" }, { "id": "p-bundle", "label": "Bundle Publisher", "kind": "integration", "sub": "signed · immutable" } ] } ] }, { "title": "Execution Plane · what a run actually does", "kind": "cloud", "icon": "azure", "dir": "row", "children": [ { "title": "Orchestration", "kind": "plain", "nodes": [ { "id": "e-orch", "label": "Durable Orchestrator", "kind": "app", "sub": "replay-safe", "icon": "azure_functions" }, { "id": "e-sched", "label": "Scheduler", "kind": "app", "sub": "queue classes" }, { "id": "e-bus", "label": "Work Queues", "kind": "queue", "sub": "Service Bus", "icon": "azure_azure-service-bus" } ] }, { "title": "Runtime", "kind": "plain", "nodes": [ { "id": "e-work", "label": "Runtime Workers", "kind": "app", "sub": "stateless · KEDA", "icon": "azure_worker-container-app" }, { "id": "e-side", "label": "Policy Sidecar", "kind": "security", "sub": "local decisions", "icon": "opa" }, { "id": "e-sbx", "label": "Session Sandbox", "kind": "platform", "sub": "Hyper-V isolated", "icon": "azure_azure-enclaves" } ] }, { "title": "Execution data", "kind": "plain", "nodes": [ { "id": "e-state", "label": "Execution Store", "kind": "store", "sub": "Cosmos · partitioned", "icon": "azure_azure-cosmos-db" }, { "id": "e-cache", "label": "Turn Cache", "kind": "store", "sub": "Managed Redis", "icon": "azure_redis" }, { "id": "e-art", "label": "Artifact Store", "kind": "store", "sub": "ADLS Gen2", "icon": "azure_datalake" } ] } ] } ], "outside": [ { "id": "o-gw", "label": "AI Gateway and Tool Broker", "kind": "integration", "sub": "APIM · shared" }, { "id": "o-obs", "label": "Observability and Cost Plane", "kind": "platform", "sub": "Monitor · ADX" } ], "edges": [ { "from": "p-bundle", "to": "e-work", "label": "signed bundle", "kind": "async" }, { "from": "p-sec", "to": "e-work", "label": "scoped token" }, { "from": "e-orch", "to": "e-state", "label": "checkpoint" }, { "from": "e-orch", "to": "e-bus", "label": "dispatch", "kind": "async" }, { "from": "e-work", "to": "o-gw", "label": "gateway" }, { "from": "e-work", "to": "o-obs", "label": "telemetry", "kind": "async" } ], "note": "A control-plane outage stops new admissions. Runs already admitted carry a pinned bundle and continue.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "04-layered-architecture", "title": "Layered Architecture — What Depends on What", "layout": "bands", "canvas": { "width": 1800 }, "layerHeaderWidth": 160, "bands": [ { "name": "Experience", "nodes": [ { "id": "l-builder", "label": "Agent Builder", "kind": "app" }, { "id": "l-designer", "label": "Workflow Designer", "kind": "app" }, { "id": "l-console", "label": "Admin Console", "kind": "app" }, { "id": "l-sdk", "label": "SDK and CLI", "kind": "app", "icon": "python" }, { "id": "l-chan", "label": "Chat and App Channels", "kind": "app" } ] }, { "name": "Platform API", "nodes": [ { "id": "l-apim", "label": "API Management", "kind": "integration", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "l-authz", "label": "AuthN and RBAC", "kind": "security", "icon": "azure_aad" }, { "id": "l-stream", "label": "Streaming Endpoint", "kind": "integration", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "l-hook", "label": "Webhook Ingress", "kind": "integration", "icon": "azure_AzureAPIManagement (Web)" } ] }, { "name": "Control plane", "nodes": [ { "id": "l-areg", "label": "Agent Registry", "kind": "app" }, { "id": "l-wreg", "label": "Workflow Registry", "kind": "app" }, { "id": "l-treg", "label": "Tool Registry", "kind": "app", "icon": "azure_api-center" }, { "id": "l-mreg", "label": "Model Registry", "kind": "app", "icon": "azure_foundry-control-plane" }, { "id": "l-polsvc", "label": "Policy Service", "kind": "security", "icon": "opa" }, { "id": "l-prompt", "label": "Prompt and Config", "kind": "app" } ] }, { "name": "Orchestration", "nodes": [ { "id": "l-orch", "label": "Durable Orchestrator", "kind": "app", "icon": "azure_functions" }, { "id": "l-sched", "label": "Scheduler", "kind": "app" }, { "id": "l-coord", "label": "Agent Coordinator", "kind": "app" }, { "id": "l-evt", "label": "Event Manager", "kind": "app" }, { "id": "l-appr", "label": "Approval Service", "kind": "app", "icon": "fa5_user_check" }, { "id": "l-statemgr", "label": "State Manager", "kind": "app", "icon": "fa5_cube" } ] }, { "name": "Agent runtime", "nodes": [ { "id": "l-exec", "label": "Agent Executor", "kind": "app", "icon": "agent" }, { "id": "l-ctx", "label": "Context Manager", "kind": "app" }, { "id": "l-guard", "label": "Guardrail Chain", "kind": "security", "icon": "azure_content-safety" }, { "id": "l-toolx", "label": "Tool Executor", "kind": "app" }, { "id": "l-mem", "label": "Memory Service", "kind": "app", "icon": "fa5_memory" }, { "id": "l-ckpt", "label": "Checkpointer", "kind": "app" }, { "id": "l-sbx", "label": "Sandbox", "kind": "platform" } ] }, { "name": "Access", "nodes": [ { "id": "l-gw", "label": "AI Gateway", "kind": "integration" }, { "id": "l-broker", "label": "Tool Broker", "kind": "integration" }, { "id": "l-ret", "label": "Retrieval Service", "kind": "app" }, { "id": "l-credb", "label": "Credential Broker", "kind": "security", "icon": "azure_AzureKeyVault (Security)" } ] }, { "name": "Data", "nodes": [ { "id": "l-sql", "label": "Azure SQL", "kind": "store", "icon": "azure_azure-sql" }, { "id": "l-cos", "label": "Cosmos DB", "kind": "store", "icon": "azure_azure-cosmos-db" }, { "id": "l-redis", "label": "Managed Redis", "kind": "store", "icon": "azure_redis" }, { "id": "l-search", "label": "Azure AI Search", "kind": "store", "icon": "azure_cognitive-search" }, { "id": "l-adls", "label": "ADLS Gen2", "kind": "store", "icon": "azure_datalake" }, { "id": "l-sb", "label": "Service Bus", "kind": "queue", "icon": "azure_azure-service-bus" }, { "id": "l-kv", "label": "Key Vault", "kind": "security", "icon": "azure_AzureKeyVault (Security)" } ] }, { "name": "Assurance", "nodes": [ { "id": "l-mon", "label": "Azure Monitor", "kind": "platform", "icon": "azure_monitor" }, { "id": "l-adx", "label": "Data Explorer", "kind": "store", "icon": "azure_azure-data-explorer-clusters" }, { "id": "l-graf", "label": "Managed Grafana", "kind": "platform", "icon": "azure_azure-managed-grafana" }, { "id": "l-eval", "label": "Evaluation Service", "kind": "app", "icon": "azure_azure-machine-learning" }, { "id": "l-audit", "label": "Audit Ledger", "kind": "store", "icon": "azure_blobstorage" }, { "id": "l-siem", "label": "Microsoft Sentinel", "kind": "security", "icon": "azure_sentinel" } ] } ], "edges": [ { "from": "l-sdk", "to": "l-apim", "label": "HTTPS" }, { "from": "l-apim", "to": "l-areg", "label": "REST" }, { "from": "l-apim", "to": "l-orch", "label": "submit run" }, { "from": "l-orch", "to": "l-exec", "label": "dispatch", "kind": "async" }, { "from": "l-exec", "to": "l-gw", "label": "inference" }, { "from": "l-exec", "to": "l-broker", "label": "tool call" }, { "from": "l-gw", "to": "l-cos", "label": "meter", "kind": "async" }, { "from": "l-ckpt", "to": "l-cos", "label": "state" }, { "from": "l-exec", "to": "l-mon", "label": "spans", "kind": "async" }, { "from": "l-eval", "to": "l-prompt", "label": "promotion gate", "kind": "batch" } ], "note": "Dependencies point downward. The one exception is drawn: evaluation scores feed the promotion gate in the control plane.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "05-container-architecture", "title": "Container and Component Architecture — Azure Deployable Units", "layout": "nested", "canvas": { "width": 1780 }, "boxes": [ { "title": "Subscription: agentplat-prod · VNet 10.40.0.0/16", "kind": "cloud", "icon": "azure", "dir": "col", "children": [ { "title": "Edge subnet", "kind": "boundary", "nodes": [ { "id": "c-fd", "label": "Front Door + WAF", "kind": "integration", "sub": "Premium · Private Link", "icon": "azure_AzureFrontDoor (Networking)" }, { "id": "c-apim", "label": "API Management", "kind": "integration", "sub": "Premium · internal VNet", "icon": "azure_AzureAPIManagement (Web)" } ] }, { "title": "Control plane subnet · AKS (3 AZs)", "kind": "boundary", "nodes": [ { "id": "c-regapi", "label": "Registry API", "kind": "app", "sub": ".NET 9 · 6 pods" }, { "id": "c-polsvc", "label": "Policy Service", "kind": "security", "sub": "OPA · 4 pods", "icon": "opa" }, { "id": "c-cfg", "label": "Config Service", "kind": "app", "sub": "App Config · 3 pods", "icon": "azure_app-configuration" }, { "id": "c-cred", "label": "Credential Broker", "kind": "security", "sub": "workload identity", "icon": "azure_AzureKeyVault (Security)" }, { "id": "c-bff", "label": "Console BFF", "kind": "app", "sub": "TypeScript · 4 pods" } ] }, { "title": "Execution plane · Container Apps environment", "kind": "boundary", "nodes": [ { "id": "c-orch", "label": "Durable Orchestrator", "kind": "app", "sub": "Functions · 10-80", "icon": "azure_functions" }, { "id": "c-sched", "label": "Scheduler", "kind": "app", "sub": "leader-elected · 3" }, { "id": "c-work", "label": "Runtime Workers", "kind": "app", "sub": "Python · KEDA 20-400", "icon": "azure_worker-container-app" }, { "id": "c-toolx", "label": "Tool Executor", "kind": "app", "sub": "Go · KEDA 10-120" }, { "id": "c-appr", "label": "Approval Service", "kind": "app", "sub": ".NET 9 · 3 pods", "icon": "fa5_user_check" }, { "id": "c-mem", "label": "Memory Service", "kind": "app", "sub": "Python · 6 pods", "icon": "fa5_memory" } ] }, { "title": "Session pool · Hyper-V isolated, no default egress", "kind": "boundary", "nodes": [ { "id": "c-code", "label": "Code Sessions", "kind": "platform", "sub": "dynamic sessions", "icon": "azure_azure-enclaves" }, { "id": "c-brow", "label": "Browser Sessions", "kind": "platform", "sub": "computer use", "icon": "azure_azure-enclaves" } ] }, { "title": "Data subnet · private endpoints only", "kind": "boundary", "nodes": [ { "id": "c-sql", "label": "Azure SQL", "kind": "store", "sub": "Business Critical", "icon": "azure_azure-sql" }, { "id": "c-cos", "label": "Cosmos DB", "kind": "store", "sub": "session consistency", "icon": "azure_azure-cosmos-db" }, { "id": "c-redis", "label": "Managed Redis", "kind": "store", "sub": "Balanced B5", "icon": "azure_redis" }, { "id": "c-search", "label": "Azure AI Search", "kind": "store", "sub": "S2 · 3 replicas", "icon": "azure_cognitive-search" }, { "id": "c-adls", "label": "ADLS Gen2", "kind": "store", "sub": "ZRS · lifecycle", "icon": "azure_datalake" }, { "id": "c-sb", "label": "Service Bus", "kind": "queue", "sub": "Premium · 4 MU", "icon": "azure_azure-service-bus" }, { "id": "c-kv", "label": "Key Vault", "kind": "security", "sub": "HSM · RBAC", "icon": "azure_AzureKeyVault (Security)" } ] } ] } ], "outside": [ { "id": "c-foundry", "label": "Azure AI Foundry", "kind": "external", "sub": "PTU + PAYG", "icon": "azure_ai-foundry" }, { "id": "c-mcp", "label": "MCP Tool Servers", "kind": "external", "sub": "internal + partner", "icon": "fa5_plug" }, { "id": "c-saas", "label": "SaaS Applications", "kind": "external", "sub": "SAP · SFDC · SNOW", "icon": "fa5_building" } ], "edges": [ { "from": "c-fd", "to": "c-apim", "label": "HTTPS" }, { "from": "c-apim", "to": "c-regapi", "label": "REST" }, { "from": "c-apim", "to": "c-orch", "label": "submit run" }, { "from": "c-orch", "to": "c-sb", "label": "dispatch", "kind": "async" }, { "from": "c-work", "to": "c-code", "label": "exec request" }, { "from": "c-work", "to": "c-apim", "label": "gateway hop" }, { "from": "c-apim", "to": "c-foundry", "label": "inference" }, { "from": "c-toolx", "to": "c-mcp", "label": "MCP" }, { "from": "c-toolx", "to": "c-saas", "label": "REST" }, { "from": "c-orch", "to": "c-cos", "label": "checkpoint" } ], "note": "Replica ranges are the sizing basis for 5,000 concurrent executions. Every data service is reached over a private endpoint.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "06-integration-catalogue", "title": "Integration and Interface Catalogue", "layout": "hub", "left": { "title": "Inbound", "nodes": [ { "id": "i-chat", "label": "Chat Channels", "kind": "external", "rel": "SSE · realtime", "icon": "teams" }, { "id": "i-app", "label": "Business Applications", "kind": "external", "rel": "REST · realtime", "icon": "fa5_building" }, { "id": "i-evt", "label": "Event Sources", "kind": "external", "rel": "CloudEvents", "kind2": "async", "icon": "azure_event-grid-topics" }, { "id": "i-hook", "label": "Partner Webhooks", "kind": "external", "rel": "HMAC signed", "kind2": "async" }, { "id": "i-cicd", "label": "CI/CD Pipelines", "kind": "external", "rel": "OIDC · deploy" } ] }, "centre": { "title": "Platform edge", "nodes": [ { "id": "i-fd", "label": "Front Door + WAF", "kind": "integration", "sub": "TLS 1.3 · anycast", "icon": "azure_AzureFrontDoor (Networking)" }, { "id": "i-apim", "label": "API Management", "kind": "integration", "sub": "38 public operations", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "i-eg", "label": "Event Grid", "kind": "queue", "sub": "6 published topics", "icon": "azure_event-grid-topics" }, { "id": "i-timer", "label": "Scheduled Triggers", "kind": "integration", "sub": "cron · per project" }, { "id": "i-core", "label": "Platform Core", "kind": "app", "sub": "control + execution" } ] }, "right": { "title": "Outbound", "nodes": [ { "id": "i-foundry", "label": "Azure AI Foundry", "kind": "external", "rel": "inference · sync", "dir": "out", "icon": "azure_ai-foundry" }, { "id": "i-fm", "label": "Foundry Models", "kind": "external", "rel": "inference · sync", "dir": "out", "icon": "azure_foundry-models" }, { "id": "i-mcp", "label": "MCP Tool Servers", "kind": "external", "rel": "MCP · on demand", "dir": "out", "icon": "fa5_plug" }, { "id": "i-api", "label": "Enterprise APIs", "kind": "external", "rel": "REST · on demand", "dir": "out", "icon": "fa5_plug" }, { "id": "i-saas", "label": "SaaS Applications", "kind": "external", "rel": "REST · on demand", "dir": "out", "icon": "fa5_building" }, { "id": "i-graph", "label": "Microsoft Graph", "kind": "external", "rel": "cards · events", "dir": "out", "kind2": "async", "icon": "teams" }, { "id": "i-purv", "label": "Microsoft Purview", "kind": "external", "rel": "lineage · nightly", "dir": "out", "kind2": "batch", "icon": "azure_AzurePurviewAccount (Analytics)" }, { "id": "i-sent", "label": "Microsoft Sentinel", "kind": "external", "rel": "security events", "dir": "out", "kind2": "async", "icon": "azure_sentinel" } ] }, "note": "Every interface is owned by the platform except the MCP wire contract and the Graph card schema.", "meta": { "v": "1.0", "owner": "Integration Architecture", "date": "2026-08" }, "canvas": { "width": 1800 } }, { "id": "07-agent-runtime-internals", "title": "Agent Runtime — Inside One Worker", "layout": "flow", "chain": true, "canvas": { "width": 1800 }, "stages": [ { "title": "Lease", "nodes": [ { "id": "r-lease", "label": "Queue Lease", "kind": "queue", "sub": "session locked" }, { "id": "r-idem", "label": "Idempotency Check", "kind": "decision", "sub": "step key" }, { "id": "r-bundle", "label": "Bundle Load", "kind": "app", "sub": "pinned version" } ] }, { "title": "Prepare", "nodes": [ { "id": "r-ctx", "label": "Context Manager", "kind": "app", "sub": "assemble window" }, { "id": "r-memr", "label": "Memory Read", "kind": "app", "sub": "scoped · audited", "icon": "fa5_memory" }, { "id": "r-limit", "label": "Limit Check", "kind": "security", "sub": "tokens · depth · cost" } ] }, { "title": "Guard in", "nodes": [ { "id": "r-shield", "label": "Prompt Shields", "kind": "security", "sub": "injection · jailbreak", "icon": "azure_content-safety" }, { "id": "r-pii", "label": "PII Redaction", "kind": "security", "sub": "detect · mask" }, { "id": "r-cls", "label": "Classification Gate", "kind": "decision", "sub": "residency route" } ] }, { "title": "Reason", "nodes": [ { "id": "r-agent", "label": "Agent Loop", "kind": "app", "sub": "framework adapter", "icon": "agent" }, { "id": "r-gwcall", "label": "AI Gateway Call", "kind": "integration", "sub": "stream · retries", "icon": "azure_ai-gateway" }, { "id": "r-plan", "label": "Plan or Delegate", "kind": "decision", "sub": "coordinator hop" } ] }, { "title": "Act", "nodes": [ { "id": "r-tval", "label": "Tool Call Validation", "kind": "security", "sub": "schema · permission" }, { "id": "r-tool", "label": "Tool Executor", "kind": "app", "sub": "timeout · retry" }, { "id": "r-sbx", "label": "Session Sandbox", "kind": "platform", "sub": "code · browser", "icon": "azure_azure-enclaves" } ] }, { "title": "Guard out", "nodes": [ { "id": "r-schema", "label": "Structured Output", "kind": "security", "sub": "schema enforced" }, { "id": "r-ground", "label": "Groundedness Check", "kind": "security", "sub": "citation bound" }, { "id": "r-esc", "label": "Escalate to Human", "kind": "decision", "sub": "on fail" } ] }, { "title": "Commit", "nodes": [ { "id": "r-ckpt", "label": "Checkpoint", "kind": "store", "sub": "durable state" }, { "id": "r-memw", "label": "Memory Write", "kind": "app", "sub": "policy TTL", "icon": "fa5_memory" }, { "id": "r-emit", "label": "Emit Telemetry", "kind": "platform", "sub": "OTel GenAI" } ] } ], "edges": [ { "from": "r-esc", "to": "r-ckpt", "label": "awaits", "kind": "async", "route": "gutter" }, { "from": "r-tool", "to": "r-agent", "label": "tool result", "route": "gutter" } ], "note": "The worker holds no long-lived secret and no state between steps. Everything durable is written at Commit.", "meta": { "v": "1.0", "owner": "Runtime Engineering", "date": "2026-08" } }, { "id": "08-model-gateway-and-routing", "title": "Model Abstraction — AI Gateway, Routing and Fallback", "layout": "nested", "canvas": { "width": 1720 }, "boxes": [ { "title": "AI Gateway · Azure API Management", "kind": "boundary", "dir": "row", "children": [ { "title": "Request policy chain", "kind": "plain", "nodes": [ { "id": "g-auth", "label": "Identity Check", "kind": "security", "sub": "agent identity", "icon": "fa5_id_badge" }, { "id": "g-tlimit", "label": "Token Rate Limit", "kind": "security", "sub": "per tenant · TPM" }, { "id": "g-cache", "label": "Semantic Cache", "kind": "store", "sub": "Redis · vector", "icon": "azure_redis" }, { "id": "g-red", "label": "Egress Guard", "kind": "security", "sub": "classification" } ] }, { "title": "Router", "kind": "plain", "nodes": [ { "id": "g-cap", "label": "Capability Match", "kind": "decision", "sub": "modality · context" }, { "id": "g-pol", "label": "Route Policy", "kind": "decision", "sub": "cost · latency · region" }, { "id": "g-lb", "label": "Weighted Pool", "kind": "integration", "sub": "load balanced" }, { "id": "g-cb", "label": "Circuit Breaker", "kind": "security", "sub": "429 · 5xx trip" } ] }, { "title": "Response policy chain", "kind": "plain", "nodes": [ { "id": "g-meter", "label": "Token Metric Emit", "kind": "platform", "sub": "per execution" }, { "id": "g-stream", "label": "Stream Relay", "kind": "integration", "sub": "SSE passthrough" }, { "id": "g-fb", "label": "Fallback Chain", "kind": "decision", "sub": "next healthy route" } ] } ] }, { "title": "Model backends", "kind": "boundary", "dir": "row", "children": [ { "title": "Azure AI Foundry", "kind": "plain", "nodes": [ { "id": "g-ptu", "label": "Reasoning PTU Pool", "kind": "external", "sub": "reserved capacity", "icon": "azure_azure-openai" }, { "id": "g-payg", "label": "PAYG Spillover", "kind": "external", "sub": "burst · 3 regions", "icon": "azure_azure-openai" }, { "id": "g-emb", "label": "Embedding Models", "kind": "external", "sub": "batch friendly", "icon": "azure_azure-openai" } ] }, { "title": "Alternate providers", "kind": "plain", "nodes": [ { "id": "g-fm", "label": "Foundry Models", "kind": "external", "sub": "partner catalogue", "icon": "azure_foundry-models" }, { "id": "g-kaito", "label": "Open-Weight on AKS", "kind": "external", "sub": "KAITO · GPU pool", "icon": "azure_kubernetes-services" }, { "id": "g-rerank", "label": "Reranker", "kind": "external", "sub": "self-hosted" } ] } ] } ], "outside": [ { "id": "g-mreg", "label": "Model Registry", "kind": "app", "sub": "control plane", "icon": "azure_foundry-control-plane" }, { "id": "g-health", "label": "Provider Health", "kind": "platform", "sub": "probe · 30 s" }, { "id": "g-cost", "label": "Cost Meter", "kind": "store", "sub": "Event Hubs · ADX", "icon": "azure_cost-analysis" } ], "edges": [ { "from": "g-mreg", "to": "g-pol", "label": "route table", "kind": "async" }, { "from": "g-health", "to": "g-cb", "label": "backend state", "kind": "async" }, { "from": "g-lb", "to": "g-ptu", "label": "primary" }, { "from": "g-fb", "to": "g-payg", "label": "on trip", "kind": "error" }, { "from": "g-fb", "to": "g-fm", "label": "on trip", "kind": "error" }, { "from": "g-meter", "to": "g-cost", "label": "tokens", "kind": "async", "route": "gutter" } ], "note": "No agent holds a provider key. The gateway presents its own managed identity to every backend.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "09-tool-plane-and-permissions", "title": "Tool Plane — Discovery, Authorisation and Invocation by Tool Class", "layout": "swimlane", "canvas": { "width": 1800 }, "laneHeaderWidth": 150, "stages": [ "Discover", "Authorise", "Invoke", "Validate", "Record" ], "lanes": [ { "title": "HTTP and REST", "cells": [ [ { "id": "t1a", "label": "API Center entry", "kind": "app", "icon": "azure_api-center" } ], [ { "id": "t1b", "label": "Agent scope grant", "kind": "security" } ], [ { "id": "t1c", "label": "APIM egress", "kind": "integration", "icon": "azure_AzureAPIManagement (Web)" } ], [ { "id": "t1d", "label": "JSON Schema", "kind": "security" } ], [ { "id": "t1e", "label": "Tool call span", "kind": "platform" } ] ] }, { "title": "MCP servers", "cells": [ [ { "id": "t2a", "label": "MCP tools/list", "kind": "app" } ], [ { "id": "t2b", "label": "Server allow-list", "kind": "security" } ], [ { "id": "t2c", "label": "MCP over HTTP", "kind": "integration" } ], [ { "id": "t2d", "label": "Result schema", "kind": "security" } ], [ { "id": "t2e", "label": "Tool call span", "kind": "platform" } ] ] }, { "title": "Data and search", "cells": [ [ { "id": "t3a", "label": "Index catalogue", "kind": "app", "icon": "azure_cognitive-search" } ], [ { "id": "t3b", "label": "Row-level filter", "kind": "security" } ], [ { "id": "t3c", "label": "AI Search query", "kind": "store", "icon": "azure_cognitive-search" } ], [ { "id": "t3d", "label": "Citation binding", "kind": "security" } ], [ { "id": "t3e", "label": "Provenance record", "kind": "store" } ] ] }, { "title": "Code and browser", "cells": [ [ { "id": "t4a", "label": "Session template", "kind": "app" } ], [ { "id": "t4b", "label": "Egress policy", "kind": "security" } ], [ { "id": "t4c", "label": "Dynamic session", "kind": "platform", "icon": "azure_azure-enclaves" } ], [ { "id": "t4d", "label": "Output size cap", "kind": "security" } ], [ { "id": "t4e", "label": "Artifact to ADLS", "kind": "store", "icon": "azure_datalake" } ] ] }, { "title": "Sensitive actions", "cells": [ [ { "id": "t5a", "label": "Marked in registry", "kind": "app" } ], [ { "id": "t5b", "label": "Human approval", "kind": "decision" } ], [ { "id": "t5c", "label": "Idempotency key", "kind": "integration" } ], [ { "id": "t5d", "label": "Effect assertion", "kind": "security" } ], [ { "id": "t5e", "label": "Audit ledger", "kind": "store", "icon": "azure_blobstorage" } ] ] } ], "note": "Every lane ends in a record. A tool the registry does not list cannot be called, whatever the model asks for.", "meta": { "v": "1.0", "owner": "Integration Architecture", "date": "2026-08" } }, { "id": "10-messaging-and-queue-topology", "title": "Event and Messaging Topology", "layout": "nested", "canvas": { "width": 1740 }, "boxes": [ { "title": "Azure Service Bus Premium · work distribution", "kind": "boundary", "dir": "row", "children": [ { "title": "Work queue classes", "kind": "plain", "nodes": [ { "id": "q-int", "label": "Interactive", "kind": "queue", "sub": "p95 lease under 2 s", "icon": "azure_AzureServiceBusQueue (Integration)" }, { "id": "q-batch", "label": "Batch", "kind": "queue", "sub": "throughput first", "icon": "azure_AzureServiceBusQueue (Integration)" }, { "id": "q-long", "label": "Long-Running", "kind": "queue", "sub": "hours to days", "icon": "azure_AzureServiceBusQueue (Integration)" }, { "id": "q-retry", "label": "Retry Scheduled", "kind": "queue", "sub": "backoff + jitter", "icon": "azure_AzureServiceBusQueue (Integration)" } ] }, { "title": "Delivery controls", "kind": "plain", "nodes": [ { "id": "q-sess", "label": "Session Ordering", "kind": "integration", "sub": "per execution" }, { "id": "q-lock", "label": "Lock Renewal", "kind": "integration", "sub": "heartbeat", "icon": "fa5_heartbeat" }, { "id": "q-dlq", "label": "Dead Letter", "kind": "queue", "sub": "per queue", "icon": "azure_AzureServiceBusQueue (Integration)" } ] } ] }, { "title": "Azure Event Grid · lifecycle and webhooks", "kind": "boundary", "dir": "row", "children": [ { "title": "Published topics", "kind": "plain", "nodes": [ { "id": "q-agt", "label": "Agent Lifecycle", "kind": "queue", "sub": "start · pause · end", "icon": "azure_event-grid-topics" }, { "id": "q-wf", "label": "Workflow Events", "kind": "queue", "sub": "state transitions", "icon": "azure_event-grid-topics" }, { "id": "q-appr", "label": "Approval Events", "kind": "queue", "sub": "raised · decided", "icon": "azure_event-grid-topics" }, { "id": "q-guard", "label": "Guardrail Events", "kind": "queue", "sub": "block · escalate", "icon": "azure_event-grid-topics" } ] }, { "title": "External edges", "kind": "plain", "nodes": [ { "id": "q-in", "label": "Webhook Ingress", "kind": "integration", "sub": "HMAC · replay guard", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "q-out", "label": "Partner Delivery", "kind": "integration", "sub": "retry 24 h" } ] } ] }, { "title": "Azure Event Hubs · telemetry firehose", "kind": "boundary", "dir": "row", "children": [ { "title": "Streams", "kind": "plain", "nodes": [ { "id": "q-otel", "label": "Trace Stream", "kind": "queue", "sub": "OTel GenAI", "icon": "azure_event-hubs" }, { "id": "q-tok", "label": "Token Meter", "kind": "queue", "sub": "per model call", "icon": "azure_event-hubs" }, { "id": "q-cap", "label": "Capture to ADLS", "kind": "store", "sub": "Avro · 5 min", "icon": "azure_datalake" } ] } ] } ], "outside": [ { "id": "q-triage", "label": "DLQ Triage Worker", "kind": "app", "sub": "classify · replay" }, { "id": "q-alert", "label": "Operator Alert", "kind": "platform", "sub": "depth and age SLO" } ], "edges": [ { "from": "q-dlq", "to": "q-triage", "label": "poison message", "kind": "error" }, { "from": "q-triage", "to": "q-retry", "label": "replay", "kind": "async" }, { "from": "q-dlq", "to": "q-alert", "label": "depth breach", "kind": "error" }, { "from": "q-appr", "to": "q-out", "label": "notify", "kind": "async" }, { "from": "q-tok", "to": "q-cap", "label": "capture", "kind": "batch" } ], "note": "Delivery is at-least-once everywhere. Effects are made idempotent at the handler, never assumed at the broker.", "meta": { "v": "1.0", "owner": "Platform Engineering", "date": "2026-08" } }, { "id": "11-definition-lifecycle", "title": "Definition Lifecycle — Versioning an Agent, Workflow, Prompt or Policy", "layout": "flow", "chain": true, "canvas": { "width": 1760 }, "stages": [ { "title": "Author", "nodes": [ { "id": "d-spec", "label": "Agent Spec", "kind": "app", "sub": "YAML · typed schema" }, { "id": "d-local", "label": "Local Runtime", "kind": "app", "sub": "same contract", "icon": "python" }, { "id": "d-git", "label": "Git Repository", "kind": "store", "sub": "source of truth", "icon": "github" } ] }, { "title": "Validate", "nodes": [ { "id": "d-lint", "label": "Schema Lint", "kind": "security", "sub": "spec version" }, { "id": "d-policy", "label": "Policy Simulation", "kind": "security", "sub": "dry run", "icon": "opa" }, { "id": "d-eval", "label": "Golden Evaluation", "kind": "app", "sub": "regression suite", "icon": "azure_azure-machine-learning" } ] }, { "title": "Review", "nodes": [ { "id": "d-diff", "label": "Behaviour Diff", "kind": "app", "sub": "vs live version" }, { "id": "d-appr", "label": "Owner Approval", "kind": "decision", "sub": "RBAC gated" }, { "id": "d-risk", "label": "Risk Classification", "kind": "security", "sub": "tool blast radius" } ] }, { "title": "Publish", "nodes": [ { "id": "d-ver", "label": "Immutable Version", "kind": "store", "sub": "content hash" }, { "id": "d-sign", "label": "Bundle Signing", "kind": "security", "sub": "Key Vault key", "icon": "azure_AzureKeyVault (Security)" }, { "id": "d-reg", "label": "Registry Entry", "kind": "app", "sub": "draft to published" } ] }, { "title": "Promote", "nodes": [ { "id": "d-dev", "label": "Development", "kind": "app", "sub": "auto", "icon": "fa5_layer_group" }, { "id": "d-test", "label": "Test and Staging", "kind": "app", "sub": "eval gate", "icon": "fa5_layer_group" }, { "id": "d-prod", "label": "Production Canary", "kind": "app", "sub": "5 percent · 24 h", "icon": "fa5_layer_group" } ] }, { "title": "Operate", "nodes": [ { "id": "d-pin", "label": "Version Pinning", "kind": "security", "sub": "at admission", "icon": "fa5_thumbtack" }, { "id": "d-ab", "label": "A/B Split", "kind": "decision", "sub": "traffic weight" }, { "id": "d-roll", "label": "Rollback", "kind": "app", "sub": "prior hash" }, { "id": "d-ret", "label": "Retire", "kind": "app", "sub": "drain in-flight" } ] } ], "edges": [ { "from": "d-eval", "to": "d-diff", "label": "scores" }, { "from": "d-prod", "to": "d-roll", "label": "canary fails", "kind": "error", "route": "gutter" }, { "from": "d-pin", "to": "d-ver", "label": "resolve hash", "route": "gutter" } ], "note": "A published version is immutable. Change means a new version; rollback means re-pointing the alias, never editing.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "12-data-architecture-zones", "title": "Data Architecture — Storage Zones by Ownership and Rebuildability", "layout": "nested", "canvas": { "width": 1740 }, "boxes": [ { "title": "Authoritative · loss is unrecoverable · backup and DR mandatory", "kind": "boundary", "dir": "row", "children": [ { "title": "Configuration and governance", "kind": "plain", "nodes": [ { "id": "z-sql", "label": "Definition Store", "kind": "store", "sub": "Azure SQL · PITR 35 d", "icon": "azure_azure-sql" }, { "id": "z-kv", "label": "Credential Store", "kind": "security", "sub": "Key Vault · soft delete", "icon": "azure_AzureKeyVault (Security)" }, { "id": "z-pol", "label": "Policy Bundles", "kind": "store", "sub": "signed · versioned", "icon": "opa" } ] }, { "title": "Execution truth", "kind": "plain", "nodes": [ { "id": "z-cos", "label": "Execution Store", "kind": "store", "sub": "Cosmos · continuous BU", "icon": "azure_azure-cosmos-db" }, { "id": "z-hist", "label": "Orchestration History", "kind": "store", "sub": "durable · replayable", "icon": "azure_azure-cosmos-db" }, { "id": "z-aud", "label": "Audit Ledger", "kind": "store", "sub": "immutable blob · 7 y", "icon": "azure_blobstorage" } ] } ] }, { "title": "Rebuildable · derived, restorable by reprocessing", "kind": "boundary", "dir": "row", "children": [ { "title": "Indexes and caches", "kind": "plain", "nodes": [ { "id": "z-search", "label": "Knowledge Index", "kind": "store", "sub": "AI Search · reindex", "icon": "azure_cognitive-search" }, { "id": "z-redis", "label": "Turn and Route Cache", "kind": "store", "sub": "Redis · TTL 60 m", "icon": "azure_redis" }, { "id": "z-semc", "label": "Semantic Cache", "kind": "store", "sub": "Redis vector · 24 h", "icon": "azure_redis" } ] }, { "title": "Analytical projections", "kind": "plain", "nodes": [ { "id": "z-adx", "label": "Telemetry Store", "kind": "store", "sub": "ADX · 90 d hot", "icon": "azure_azure-data-explorer-clusters" }, { "id": "z-proj", "label": "Run Projections", "kind": "store", "sub": "console read model", "icon": "azure_azure-cosmos-db" }, { "id": "z-cost", "label": "Cost Facts", "kind": "store", "sub": "daily rollup · 24 m", "icon": "azure_azure-data-explorer-clusters" } ] } ] }, { "title": "Large payload · lifecycle and retention policy driven", "kind": "boundary", "dir": "row", "children": [ { "title": "Object storage", "kind": "plain", "nodes": [ { "id": "z-art", "label": "Run Artifacts", "kind": "store", "sub": "ADLS · tier at 30 d", "icon": "azure_datalake" }, { "id": "z-pr", "label": "Prompt and Response", "kind": "store", "sub": "opt-in · 30 d", "icon": "azure_datalake" }, { "id": "z-eval", "label": "Golden Datasets", "kind": "store", "sub": "versioned · retained", "icon": "azure_datalake" }, { "id": "z-cap", "label": "Event Capture", "kind": "store", "sub": "Avro · 12 m", "icon": "azure_datalake" } ] } ] } ], "outside": [ { "id": "z-vault", "label": "Backup Vault", "kind": "platform", "sub": "GRS · immutable", "icon": "azure_backup-vault" }, { "id": "z-geo", "label": "Paired-Region Replica", "kind": "store", "sub": "RPO 5 min", "icon": "azure_resiliency" }, { "id": "z-purv", "label": "Microsoft Purview", "kind": "external", "sub": "catalogue · lineage", "icon": "azure_AzurePurviewAccount (Analytics)" } ], "edges": [ { "from": "z-sql", "to": "z-vault", "label": "nightly", "kind": "batch" }, { "from": "z-cos", "to": "z-geo", "label": "continuous", "kind": "async" }, { "from": "z-aud", "to": "z-vault", "label": "WORM copy", "kind": "batch" }, { "from": "z-art", "to": "z-search", "label": "reindex", "kind": "batch" }, { "from": "z-cap", "to": "z-adx", "label": "ingest", "kind": "batch" }, { "from": "z-sql", "to": "z-purv", "label": "scan", "kind": "batch" } ], "note": "Only the top zone needs disaster recovery. Everything in the middle zone is rebuilt from it, and that is the sizing basis for RTO.", "meta": { "v": "1.0", "owner": "Data Architecture", "date": "2026-08" } }, { "id": "13-memory-subsystem", "title": "Memory Subsystem — Scope, Store, Retrieval, Permission, Retention", "layout": "grid", "canvas": { "width": 1800 }, "laneHeaderWidth": 160, "columns": [ "Written by", "Store", "Retrieval", "Permission", "Retention" ], "rows": [ { "title": "Execution state", "cells": [ [ { "id": "y1a", "label": "Orchestrator", "kind": "app" } ], [ { "id": "y1b", "label": "Cosmos DB", "kind": "store", "icon": "azure_azure-cosmos-db" } ], [ { "id": "y1c", "label": "By execution id", "kind": "integration" } ], [ { "id": "y1d", "label": "Run owner only", "kind": "security" } ], [ { "id": "y1e", "label": "90 days", "kind": "platform" } ] ] }, { "title": "Conversation", "cells": [ [ { "id": "y2a", "label": "Agent Executor", "kind": "app", "icon": "agent" } ], [ { "id": "y2b", "label": "Cosmos DB", "kind": "store", "icon": "azure_azure-cosmos-db" } ], [ { "id": "y2c", "label": "Recent turns", "kind": "integration" } ], [ { "id": "y2d", "label": "Session subject", "kind": "security" } ], [ { "id": "y2e", "label": "Tenant policy", "kind": "platform" } ] ] }, { "title": "Episodic", "cells": [ [ { "id": "y3a", "label": "Memory Service", "kind": "app", "icon": "fa5_memory" } ], [ { "id": "y3b", "label": "Cosmos + vectors", "kind": "store", "icon": "azure_azure-cosmos-db" } ], [ { "id": "y3c", "label": "Similarity + recency", "kind": "integration" } ], [ { "id": "y3d", "label": "Agent scope grant", "kind": "security" } ], [ { "id": "y3e", "label": "Decay then expire", "kind": "platform" } ] ] }, { "title": "Knowledge", "cells": [ [ { "id": "y4a", "label": "Ingestion Pipeline", "kind": "app" } ], [ { "id": "y4b", "label": "Azure AI Search", "kind": "store", "icon": "azure_cognitive-search" } ], [ { "id": "y4c", "label": "Hybrid + reranker", "kind": "integration" } ], [ { "id": "y4d", "label": "Security trimming", "kind": "security" } ], [ { "id": "y4e", "label": "Source lifecycle", "kind": "platform" } ] ] }, { "title": "Profile and org", "cells": [ [ { "id": "y5a", "label": "Admin and user", "kind": "actor" } ], [ { "id": "y5b", "label": "Azure SQL", "kind": "store", "icon": "azure_azure-sql" } ], [ { "id": "y5c", "label": "Direct lookup", "kind": "integration" } ], [ { "id": "y5d", "label": "Consent flagged", "kind": "security" } ], [ { "id": "y5e", "label": "Erasure on request", "kind": "platform" } ] ] }, { "title": "Artifacts", "cells": [ [ { "id": "y6a", "label": "Tool Executor", "kind": "app" } ], [ { "id": "y6b", "label": "ADLS Gen2", "kind": "store", "icon": "azure_datalake" } ], [ { "id": "y6c", "label": "By artifact id", "kind": "integration" } ], [ { "id": "y6d", "label": "SAS · 15 minutes", "kind": "security" } ], [ { "id": "y6e", "label": "Tier then delete", "kind": "platform" } ] ] } ], "note": "Every read is filtered by tenant and by the agent's declared memory scope, and every read is written to the audit ledger.", "meta": { "v": "1.0", "owner": "Data Architecture", "date": "2026-08" } }, { "id": "14-context-assembly-pipeline", "title": "Context Management — Assembling One Model Invocation", "layout": "flow", "chain": true, "canvas": { "width": 1780 }, "stages": [ { "title": "Resolve", "nodes": [ { "id": "x-strat", "label": "Context Strategy", "kind": "app", "sub": "per agent spec" }, { "id": "x-budget", "label": "Window Budget", "kind": "decision", "sub": "reserve for output" } ] }, { "title": "Gather", "nodes": [ { "id": "x-sys", "label": "System Instruction", "kind": "app", "sub": "pinned version" }, { "id": "x-hist", "label": "Conversation Slice", "kind": "store", "sub": "last N turns" }, { "id": "x-mem", "label": "Memory Recall", "kind": "app", "sub": "scoped query", "icon": "fa5_memory" }, { "id": "x-kb", "label": "Knowledge Retrieval", "kind": "store", "sub": "hybrid search", "icon": "azure_cognitive-search" } ] }, { "title": "Rank", "nodes": [ { "id": "x-rerank", "label": "Reranker", "kind": "app", "sub": "cross-encoder" }, { "id": "x-dedup", "label": "Deduplicate", "kind": "app", "sub": "near-duplicate drop" }, { "id": "x-rel", "label": "Relevance Cut", "kind": "decision", "sub": "score threshold" } ] }, { "title": "Compress", "nodes": [ { "id": "x-sum", "label": "Rolling Summary", "kind": "app", "sub": "cheap model" }, { "id": "x-trim", "label": "Tool Output Trim", "kind": "app", "sub": "head and tail" }, { "id": "x-fit", "label": "Fit to Window", "kind": "decision", "sub": "hard cap" } ] }, { "title": "Protect", "nodes": [ { "id": "x-pii", "label": "PII Redaction", "kind": "security", "sub": "reversible token" }, { "id": "x-untrust", "label": "Untrusted Fencing", "kind": "security", "sub": "content delimiters" }, { "id": "x-class", "label": "Classification Check", "kind": "security", "sub": "route constraint" } ] }, { "title": "Assemble", "nodes": [ { "id": "x-msg", "label": "Message Array", "kind": "app", "sub": "provider neutral" }, { "id": "x-prov", "label": "Provenance Tags", "kind": "app", "sub": "source per block" }, { "id": "x-acct", "label": "Token Accounting", "kind": "platform", "sub": "pre-call estimate" } ] } ], "edges": [ { "from": "x-fit", "to": "x-rel", "label": "still over budget", "kind": "error", "route": "gutter" }, { "from": "x-class", "to": "x-msg", "label": "route pinned" }, { "from": "x-prov", "to": "x-acct", "label": "cited blocks" } ], "note": "Retrieved content is fenced as untrusted before it reaches the model. Provenance survives into the trace, so any claim is attributable.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "15-data-flow-by-class", "title": "Data Flow by Data Class — Produce, Transport, Land, Serve, Retain", "layout": "swimlane", "canvas": { "width": 1800 }, "laneHeaderWidth": 165, "stages": [ "Produce", "Transport", "Land", "Serve", "Retain" ], "lanes": [ { "title": "Definitions", "cells": [ [ { "id": "f1a", "label": "Git commit", "kind": "app" } ], [ { "id": "f1b", "label": "Pipeline · on merge", "kind": "integration" } ], [ { "id": "f1c", "label": "Azure SQL", "kind": "store", "icon": "azure_azure-sql" } ], [ { "id": "f1d", "label": "Signed bundle", "kind": "integration" } ], [ { "id": "f1e", "label": "All versions kept", "kind": "platform" } ] ] }, { "title": "Execution state", "cells": [ [ { "id": "f2a", "label": "Orchestrator step", "kind": "app" } ], [ { "id": "f2b", "label": "Checkpoint · sync", "kind": "integration" } ], [ { "id": "f2c", "label": "Cosmos DB", "kind": "store", "icon": "azure_azure-cosmos-db" } ], [ { "id": "f2d", "label": "Query and replay", "kind": "integration" } ], [ { "id": "f2e", "label": "90 d then archive", "kind": "platform" } ] ] }, { "title": "Prompts and output", "cells": [ [ { "id": "f3a", "label": "Model call", "kind": "app" } ], [ { "id": "f3b", "label": "Gateway capture", "kind": "integration", "icon": "azure_ai-gateway" } ], [ { "id": "f3c", "label": "ADLS · encrypted", "kind": "store", "icon": "azure_datalake" } ], [ { "id": "f3d", "label": "Trace deep link", "kind": "integration" } ], [ { "id": "f3e", "label": "Off by default", "kind": "platform" } ] ] }, { "title": "Memory and knowledge", "cells": [ [ { "id": "f4a", "label": "Source system", "kind": "external" } ], [ { "id": "f4b", "label": "Ingestion · nightly", "kind": "integration" } ], [ { "id": "f4c", "label": "AI Search index", "kind": "store", "icon": "azure_cognitive-search" } ], [ { "id": "f4d", "label": "Trimmed retrieval", "kind": "integration" } ], [ { "id": "f4e", "label": "Follows source", "kind": "platform" } ] ] }, { "title": "Telemetry and cost", "cells": [ [ { "id": "f5a", "label": "OTel span", "kind": "app" } ], [ { "id": "f5b", "label": "Event Hubs", "kind": "queue", "icon": "azure_event-hubs" } ], [ { "id": "f5c", "label": "Monitor + ADX", "kind": "store" } ], [ { "id": "f5d", "label": "Grafana · console", "kind": "platform", "icon": "azure_azure-managed-grafana" } ], [ { "id": "f5e", "label": "90 d hot · 2 y cold", "kind": "platform" } ] ] }, { "title": "Audit", "cells": [ [ { "id": "f6a", "label": "Privileged action", "kind": "security" } ], [ { "id": "f6b", "label": "Append only", "kind": "integration" } ], [ { "id": "f6c", "label": "Immutable blob", "kind": "store", "icon": "azure_blobstorage" } ], [ { "id": "f6d", "label": "Sentinel + export", "kind": "security", "icon": "azure_sentinel" } ], [ { "id": "f6e", "label": "7 years · WORM", "kind": "platform" } ] ] } ], "note": "Prompt and completion capture is opt-in per project and defaults off. Everything else on this page is always on.", "meta": { "v": "1.0", "owner": "Data Architecture", "date": "2026-08" } }, { "id": "16-platform-data-model", "title": "Platform Data Model — Core Entities", "layout": "er", "canvas": { "width": 1660, "cols": 4 }, "rowGap": 260, "entities": [ { "id": "m-ten", "name": "tenant", "kind": "store", "row": 0, "col": 0, "attrs": [ "tenant_id PK", "entra_tenant UQ", "residency_region", "quota_profile", "budget_ceiling" ] }, { "id": "m-prj", "name": "project", "kind": "store", "row": 0, "col": 1, "attrs": [ "project_id PK", "tenant_id FK", "environment", "monthly_budget", "capture_prompts" ] }, { "id": "m-agv", "name": "agent_version", "kind": "store", "row": 0, "col": 2, "attrs": [ "agent_version_id PK", "project_id FK", "spec_hash UQ", "state", "model_route_id FK" ] }, { "id": "m-wfv", "name": "workflow_version", "kind": "store", "row": 0, "col": 3, "attrs": [ "workflow_version_id PK", "graph_hash UQ", "pattern", "state", "compat_range" ] }, { "id": "m-pol", "name": "policy", "kind": "store", "row": 1, "col": 0, "attrs": [ "policy_id PK", "scope_level", "scope_ref", "bundle_hash", "effective_from" ] }, { "id": "m-exe", "name": "execution", "kind": "store", "row": 1, "col": 1, "attrs": [ "execution_id PK", "project_id FK", "workflow_version_id FK", "status", "correlation_id" ] }, { "id": "m-stp", "name": "step", "kind": "store", "row": 1, "col": 2, "attrs": [ "step_id PK", "execution_id FK", "agent_version_id FK", "attempt", "checkpoint_ref" ] }, { "id": "m-tb", "name": "tool_binding", "kind": "store", "row": 1, "col": 3, "attrs": [ "binding_id PK", "tool_version_id", "agent_version_id FK", "sensitivity", "credential_ref" ] }, { "id": "m-mem", "name": "memory_item", "kind": "store", "row": 2, "col": 0, "attrs": [ "memory_id PK", "scope_type", "scope_ref", "embedding_ref", "expires_at" ] }, { "id": "m-cost", "name": "cost_record", "kind": "store", "row": 2, "col": 1, "attrs": [ "cost_id PK", "execution_id FK", "prompt_tokens", "completion_tokens", "amount_usd" ] }, { "id": "m-tc", "name": "tool_call", "kind": "store", "row": 2, "col": 2, "attrs": [ "tool_call_id PK", "step_id FK", "idempotency_key UQ", "outcome", "duration_ms" ] }, { "id": "m-apr", "name": "approval", "kind": "store", "row": 2, "col": 3, "attrs": [ "approval_id PK", "binding_id FK", "step_id FK", "decision", "decided_by" ] } ], "relations": [ { "from": "m-ten", "to": "m-prj", "label": "1 : N", "from_side": "e", "to_side": "w" }, { "from": "m-ten", "to": "m-pol", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "m-prj", "to": "m-agv", "label": "1 : N", "from_side": "e", "to_side": "w" }, { "from": "m-agv", "to": "m-wfv", "label": "N : M", "from_side": "e", "to_side": "w" }, { "from": "m-prj", "to": "m-exe", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "m-exe", "to": "m-stp", "label": "1 : N", "from_side": "e", "to_side": "w" }, { "from": "m-agv", "to": "m-stp", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "m-stp", "to": "m-tc", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "m-exe", "to": "m-cost", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "m-tb", "to": "m-apr", "label": "1 : N", "from_side": "s", "to_side": "n" }, { "from": "m-pol", "to": "m-mem", "label": "governs", "from_side": "s", "to_side": "n", "kind": "optional" } ], "note": "tenant_id is the partition key on every execution-plane container. No entity is reachable without it.", "meta": { "v": "1.0", "owner": "Data Architecture", "date": "2026-08" } }, { "id": "17-agent-execution-sequence", "title": "Execute an Agent — Critical Flow with Streaming", "layout": "sequence", "lifelines": [ { "id": "s-user", "label": "Caller", "kind": "actor" }, { "id": "s-apim", "label": "API Management", "kind": "integration", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "s-adm", "label": "Admission Service", "kind": "app" }, { "id": "s-orch", "label": "Durable Orchestrator", "kind": "app", "icon": "azure_functions" }, { "id": "s-work", "label": "Runtime Worker", "kind": "app", "icon": "azure_worker-container-app" }, { "id": "s-gw", "label": "AI Gateway", "kind": "integration" }, { "id": "s-tool", "label": "Tool Broker", "kind": "integration" }, { "id": "s-cos", "label": "Execution Store", "kind": "store", "icon": "azure_azure-cosmos-db" } ], "messages": [ { "from": "s-user", "to": "s-apim", "label": "POST /v1/executions", "kind": "call" }, { "from": "s-apim", "to": "s-apim", "label": "verify token, scope", "kind": "self" }, { "from": "s-apim", "to": "s-adm", "label": "submit run", "kind": "call" }, { "from": "s-adm", "to": "s-adm", "label": "pin versions, evaluate policy", "kind": "self" }, { "from": "s-adm", "to": "s-user", "label": "429 budget exhausted", "kind": "error" }, { "from": "s-adm", "to": "s-orch", "label": "start orchestration", "kind": "call" }, { "from": "s-orch", "to": "s-cos", "label": "checkpoint: admitted", "kind": "call" }, { "from": "s-orch", "to": "s-work", "label": "dispatch step", "kind": "async" }, { "from": "s-work", "to": "s-work", "label": "assemble context, guard input", "kind": "self" }, { "from": "s-work", "to": "s-gw", "label": "chat completion, stream", "kind": "call" }, { "from": "s-gw", "to": "s-user", "label": "token stream", "kind": "async" }, { "from": "s-gw", "to": "s-work", "label": "tool call requested", "kind": "return" }, { "from": "s-work", "to": "s-tool", "label": "invoke with idempotency key", "kind": "call" }, { "from": "s-tool", "to": "s-work", "label": "tool result", "kind": "return" }, { "from": "s-work", "to": "s-gw", "label": "continue with result", "kind": "call" }, { "from": "s-gw", "to": "s-work", "label": "final answer", "kind": "return" }, { "from": "s-work", "to": "s-work", "label": "guard output, enforce schema", "kind": "self" }, { "from": "s-work", "to": "s-cos", "label": "checkpoint: completed", "kind": "call" }, { "from": "s-orch", "to": "s-user", "label": "200 result and cost", "kind": "return" } ], "note": "Tokens stream from the gateway to the caller while the run is still open. The result is not final until the checkpoint commits.", "meta": { "v": "1.0", "owner": "Runtime Engineering", "date": "2026-08" } }, { "id": "18-multi-agent-patterns", "title": "Multi-Agent Coordination — Six Patterns on One Runtime", "layout": "swimlane", "canvas": { "width": 1800 }, "laneHeaderWidth": 170, "stages": [ "Trigger", "Decompose", "Delegate", "Execute", "Converge" ], "lanes": [ { "title": "Supervisor and worker", "cells": [ [ { "id": "n1a", "label": "Single goal", "kind": "app" } ], [ { "id": "n1b", "label": "Supervisor splits", "kind": "app" } ], [ { "id": "n1c", "label": "Fan-out to workers", "kind": "integration" } ], [ { "id": "n1d", "label": "Isolated contexts", "kind": "app" } ], [ { "id": "n1e", "label": "Supervisor merges", "kind": "app" } ] ] }, { "title": "Planner and executor", "cells": [ [ { "id": "n2a", "label": "Open task", "kind": "app" } ], [ { "id": "n2b", "label": "Plan as DAG", "kind": "decision" } ], [ { "id": "n2c", "label": "Steps to queue", "kind": "queue" } ], [ { "id": "n2d", "label": "Executor per step", "kind": "app" } ], [ { "id": "n2e", "label": "Replan on failure", "kind": "decision" } ] ] }, { "title": "Hierarchical teams", "cells": [ [ { "id": "n3a", "label": "Programme goal", "kind": "app" } ], [ { "id": "n3b", "label": "Team leads", "kind": "app" } ], [ { "id": "n3c", "label": "Depth limit 3", "kind": "security" } ], [ { "id": "n3d", "label": "Team sub-runs", "kind": "app" } ], [ { "id": "n3e", "label": "Roll up by level", "kind": "app" } ] ] }, { "title": "Peer collaboration", "cells": [ [ { "id": "n4a", "label": "Shared brief", "kind": "app" } ], [ { "id": "n4b", "label": "No fixed owner", "kind": "app" } ], [ { "id": "n4c", "label": "Message passing", "kind": "queue" } ], [ { "id": "n4d", "label": "Shared scratchpad", "kind": "store" } ], [ { "id": "n4e", "label": "Quorum or timeout", "kind": "decision" } ] ] }, { "title": "Debate and critic", "cells": [ [ { "id": "n5a", "label": "Draft answer", "kind": "app" } ], [ { "id": "n5b", "label": "Critic prompt", "kind": "app" } ], [ { "id": "n5c", "label": "Round cap 3", "kind": "security" } ], [ { "id": "n5d", "label": "Revise and rescore", "kind": "app" } ], [ { "id": "n5e", "label": "Best scored wins", "kind": "decision" } ] ] }, { "title": "Specialist routing", "cells": [ [ { "id": "n6a", "label": "Ambiguous intent", "kind": "app" } ], [ { "id": "n6b", "label": "Capability match", "kind": "decision" } ], [ { "id": "n6c", "label": "Registry lookup", "kind": "app" } ], [ { "id": "n6d", "label": "One specialist", "kind": "app" } ], [ { "id": "n6e", "label": "Return or reroute", "kind": "decision" } ] ] } ], "note": "All six run on the same orchestrator and the same worker image. The pattern is a workflow definition, not a different platform.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "19-durable-execution-and-replay", "title": "Durable Execution — Checkpoint, Failure, Recovery and Replay", "layout": "flow", "chain": true, "canvas": { "width": 1780 }, "stages": [ { "title": "Progress", "nodes": [ { "id": "v-step", "label": "Step Completes", "kind": "app", "sub": "worker result" }, { "id": "v-ckpt", "label": "Checkpoint Write", "kind": "store", "sub": "before ack" }, { "id": "v-ack", "label": "Queue Ack", "kind": "queue", "sub": "lock released" } ] }, { "title": "Failure", "nodes": [ { "id": "v-proc", "label": "Process Crash", "kind": "risk", "sub": "pod evicted" }, { "id": "v-node", "label": "Node or Zone Loss", "kind": "risk", "sub": "AZ outage" }, { "id": "v-prov", "label": "Provider Outage", "kind": "risk", "sub": "model 5xx" } ] }, { "title": "Detect", "nodes": [ { "id": "v-lock", "label": "Lock Expiry", "kind": "integration", "sub": "no renewal" }, { "id": "v-hb", "label": "Heartbeat Miss", "kind": "platform", "sub": "orchestrator", "icon": "fa5_heartbeat" }, { "id": "v-class", "label": "Classify Failure", "kind": "decision", "sub": "eight classes" } ] }, { "title": "Recover", "nodes": [ { "id": "v-redel", "label": "Redeliver Message", "kind": "queue", "sub": "same session" }, { "id": "v-idem", "label": "Idempotency Guard", "kind": "security", "sub": "step key seen" }, { "id": "v-back", "label": "Scheduled Retry", "kind": "queue", "sub": "backoff + jitter" }, { "id": "v-fail", "label": "Provider Failover", "kind": "integration", "sub": "next route" } ] }, { "title": "Resume", "nodes": [ { "id": "v-load", "label": "Load Checkpoint", "kind": "store", "sub": "last committed" }, { "id": "v-pin", "label": "Pinned Definition", "kind": "security", "sub": "original hash", "icon": "fa5_thumbtack" }, { "id": "v-cont", "label": "Continue Step", "kind": "app", "sub": "no re-execution" } ] }, { "title": "Exhausted", "nodes": [ { "id": "v-dlq", "label": "Dead Letter", "kind": "queue", "sub": "after max attempts", "icon": "azure_AzureServiceBusQueue (Integration)" }, { "id": "v-comp", "label": "Compensate", "kind": "app", "sub": "reverse effects" }, { "id": "v-esc", "label": "Human Escalation", "kind": "actor", "sub": "with full trace" } ] } ], "edges": [ { "from": "v-class", "to": "v-back", "label": "transient" }, { "from": "v-class", "to": "v-fail", "label": "provider" }, { "from": "v-class", "to": "v-dlq", "label": "permanent", "kind": "error", "route": "gutter" }, { "from": "v-cont", "to": "v-step", "label": "run continues", "route": "gutter" }, { "from": "v-comp", "to": "v-esc", "label": "on partial reversal", "kind": "error" } ], "note": "The checkpoint is written before the queue acknowledgement, so a crash duplicates work rather than losing it.", "meta": { "v": "1.0", "owner": "Runtime Engineering", "date": "2026-08" } }, { "id": "20-human-in-the-loop", "title": "Human-in-the-Loop — Approving a Sensitive Action", "layout": "sequence", "lifelines": [ { "id": "a-work", "label": "Runtime Worker", "kind": "app", "icon": "azure_worker-container-app" }, { "id": "a-pol", "label": "Policy Engine", "kind": "security", "icon": "opa" }, { "id": "a-appr", "label": "Approval Service", "kind": "app", "icon": "fa5_user_check" }, { "id": "a-orch", "label": "Durable Orchestrator", "kind": "app", "icon": "azure_functions" }, { "id": "a-graph", "label": "Teams via Graph", "kind": "external", "icon": "teams" }, { "id": "a-human", "label": "Approver", "kind": "actor" }, { "id": "a-aud", "label": "Audit Ledger", "kind": "store", "icon": "azure_blobstorage" } ], "messages": [ { "from": "a-work", "to": "a-pol", "label": "may I call this tool", "kind": "call" }, { "from": "a-pol", "to": "a-work", "label": "approval required", "kind": "return" }, { "from": "a-work", "to": "a-appr", "label": "raise request with evidence", "kind": "call" }, { "from": "a-appr", "to": "a-aud", "label": "record raised", "kind": "async" }, { "from": "a-appr", "to": "a-orch", "label": "suspend, await event", "kind": "call" }, { "from": "a-appr", "to": "a-graph", "label": "adaptive card to queue", "kind": "async" }, { "from": "a-graph", "to": "a-human", "label": "notify approver role", "kind": "async" }, { "from": "a-human", "to": "a-appr", "label": "approve or reject", "kind": "call" }, { "from": "a-appr", "to": "a-appr", "label": "verify role, not requester", "kind": "self" }, { "from": "a-appr", "to": "a-aud", "label": "record decision and actor", "kind": "async" }, { "from": "a-appr", "to": "a-orch", "label": "raise external event", "kind": "call" }, { "from": "a-orch", "to": "a-work", "label": "resume from checkpoint", "kind": "call" }, { "from": "a-appr", "to": "a-orch", "label": "SLA expired, auto-reject", "kind": "error" }, { "from": "a-orch", "to": "a-aud", "label": "terminal state", "kind": "async" } ], "note": "The approver cannot be the requester. An unanswered request auto-rejects at the configured SLA and the run ends cleanly.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "21-guardrail-chain", "title": "Safety Guardrails — Detect, Decide, Act, Record by Threat Class", "layout": "swimlane", "canvas": { "width": 1720 }, "laneHeaderWidth": 180, "stages": [ "Detect", "Decide", "Act", "Record" ], "lanes": [ { "title": "Prompt injection", "cells": [ [ { "id": "k1a", "label": "Prompt Shields", "kind": "security", "icon": "azure_content-safety" } ], [ { "id": "k1b", "label": "Severity threshold", "kind": "decision" } ], [ { "id": "k1c", "label": "Strip and refuse", "kind": "app" } ], [ { "id": "k1d", "label": "Guardrail event", "kind": "queue" } ] ] }, { "title": "Sensitive data egress", "cells": [ [ { "id": "k2a", "label": "PII and label scan", "kind": "security" } ], [ { "id": "k2b", "label": "Classification rule", "kind": "decision" } ], [ { "id": "k2c", "label": "Redact or reroute", "kind": "app" } ], [ { "id": "k2d", "label": "Audit with hash", "kind": "store", "icon": "azure_blobstorage" } ] ] }, { "title": "Unsafe tool use", "cells": [ [ { "id": "k3a", "label": "Argument inspection", "kind": "security" } ], [ { "id": "k3b", "label": "Permission and blast", "kind": "decision" } ], [ { "id": "k3c", "label": "Block or escalate", "kind": "app" } ], [ { "id": "k3d", "label": "Audit ledger", "kind": "store", "icon": "azure_blobstorage" } ] ] }, { "title": "Ungrounded output", "cells": [ [ { "id": "k4a", "label": "Groundedness check", "kind": "security" } ], [ { "id": "k4b", "label": "Citation coverage", "kind": "decision" } ], [ { "id": "k4c", "label": "Retry then caveat", "kind": "app" } ], [ { "id": "k4d", "label": "Quality metric", "kind": "platform" } ] ] }, { "title": "Untrusted content", "cells": [ [ { "id": "k5a", "label": "Source trust tier", "kind": "security" } ], [ { "id": "k5b", "label": "Fence or drop", "kind": "decision" } ], [ { "id": "k5c", "label": "No tool authority", "kind": "app" } ], [ { "id": "k5d", "label": "Provenance tag", "kind": "store" } ] ] }, { "title": "Harmful content", "cells": [ [ { "id": "k6a", "label": "Content Safety", "kind": "security", "icon": "azure_content-safety" } ], [ { "id": "k6b", "label": "Category severity", "kind": "decision" } ], [ { "id": "k6c", "label": "Refuse and log", "kind": "app" } ], [ { "id": "k6d", "label": "Sentinel signal", "kind": "security", "icon": "azure_sentinel" } ] ] } ], "note": "Content retrieved by a tool never gains the caller's authority. Fencing happens before the model sees it, not after.", "meta": { "v": "1.0", "owner": "Security Architecture", "date": "2026-08" } }, { "id": "22-policy-and-budget", "title": "Policy and Budget Enforcement — Scope against Decision Point", "layout": "grid", "canvas": { "width": 1800 }, "laneHeaderWidth": 175, "columns": [ "At admission", "Before model call", "Before tool call", "Before response", "At completion" ], "rows": [ { "title": "Organisation", "cells": [ [ { "id": "b1a", "label": "Allowed regions", "kind": "security" } ], [ { "id": "b1b", "label": "Approved providers", "kind": "security" } ], [ { "id": "b1c", "label": "Banned tool classes", "kind": "security" } ], [ { "id": "b1d", "label": "Content restrictions", "kind": "security" } ], [ { "id": "b1e", "label": "Retention policy", "kind": "platform" } ] ] }, { "title": "Project", "cells": [ [ { "id": "b2a", "label": "Monthly budget", "kind": "decision", "icon": "azure_cost-budgets" } ], [ { "id": "b2b", "label": "Model allow-list", "kind": "security" } ], [ { "id": "b2c", "label": "Tool allow-list", "kind": "security" } ], [ { "id": "b2d", "label": "Capture opt-in", "kind": "platform" } ], [ { "id": "b2e", "label": "Cost attribution", "kind": "store", "icon": "azure_cost-analysis" } ] ] }, { "title": "Agent", "cells": [ [ { "id": "b3a", "label": "Run concurrency", "kind": "decision" } ], [ { "id": "b3b", "label": "Token ceiling", "kind": "security" } ], [ { "id": "b3c", "label": "Granted scopes", "kind": "security" } ], [ { "id": "b3d", "label": "Output schema", "kind": "security" } ], [ { "id": "b3e", "label": "Per-agent spend", "kind": "store", "icon": "azure_cost-budgets" } ] ] }, { "title": "Workflow", "cells": [ [ { "id": "b4a", "label": "Max duration", "kind": "decision" } ], [ { "id": "b4b", "label": "Recursion depth", "kind": "security" } ], [ { "id": "b4c", "label": "Approval steps", "kind": "decision" } ], [ { "id": "b4d", "label": "Compensation rule", "kind": "app" } ], [ { "id": "b4e", "label": "Run cost record", "kind": "store" } ] ] }, { "title": "Data classification", "cells": [ [ { "id": "b5a", "label": "Residency pinning", "kind": "security" } ], [ { "id": "b5b", "label": "Route constraint", "kind": "security" } ], [ { "id": "b5c", "label": "Egress denial", "kind": "security" } ], [ { "id": "b5d", "label": "Redaction rule", "kind": "security" } ], [ { "id": "b5e", "label": "Deletion trigger", "kind": "platform" } ] ] } ], "note": "One Rego bundle, five evaluation points. Decisions are local to the worker, so a control-plane outage cannot fail them open.", "meta": { "v": "1.0", "owner": "Governance", "date": "2026-08" } }, { "id": "23-deployment-topology", "title": "Deployment Topology — Regions, Zones and Failure Domains", "layout": "nested", "canvas": { "width": 1780 }, "boxes": [ { "title": "Global · anycast and tenant-wide services", "kind": "cloud", "icon": "azure", "dir": "row", "children": [ { "title": "Edge and identity", "kind": "plain", "nodes": [ { "id": "w-fd", "label": "Front Door Premium", "kind": "integration", "sub": "WAF · health probe", "icon": "azure_AzureFrontDoor (Networking)" }, { "id": "w-entra", "label": "Microsoft Entra ID", "kind": "security", "sub": "global tenant", "icon": "azure_aad" }, { "id": "w-dns", "label": "Private DNS Zones", "kind": "integration", "sub": "hub linked", "icon": "azure_private-link" }, { "id": "w-acr", "label": "Container Registry", "kind": "store", "sub": "geo-replicated", "icon": "azure_AzureContainerRegistry (Containers)" } ] } ] }, { "title": "Primary region · West Europe · active", "kind": "cloud", "icon": "azure", "dir": "row", "children": [ { "title": "Zones 1-3 · compute", "kind": "boundary", "nodes": [ { "id": "w-apim1", "label": "API Management", "kind": "integration", "sub": "Premium · 3 units", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "w-aks1", "label": "AKS Control Plane", "kind": "app", "sub": "3 AZ node pools", "icon": "azure_kubernetes-services" }, { "id": "w-aca1", "label": "Container Apps Env", "kind": "app", "sub": "workload profiles", "icon": "azure_container-apps-environments" }, { "id": "w-gpu1", "label": "GPU Node Pool", "kind": "app", "sub": "KAITO · spot mix", "icon": "azure_kubernetes-services" } ] }, { "title": "Zones 1-3 · state", "kind": "boundary", "nodes": [ { "id": "w-cos1", "label": "Cosmos DB", "kind": "store", "sub": "single write region", "icon": "azure_azure-cosmos-db" }, { "id": "w-sql1", "label": "Azure SQL", "kind": "store", "sub": "Business Critical", "icon": "azure_azure-sql" }, { "id": "w-sb1", "label": "Service Bus", "kind": "queue", "sub": "Premium · ZR", "icon": "azure_azure-service-bus" }, { "id": "w-redis1", "label": "Managed Redis", "kind": "store", "sub": "zone redundant", "icon": "azure_redis" }, { "id": "w-adls1", "label": "ADLS Gen2", "kind": "store", "sub": "ZRS", "icon": "azure_datalake" } ] } ] }, { "title": "Secondary region · North Europe · warm standby", "kind": "cloud", "icon": "azure", "dir": "row", "children": [ { "title": "Control plane standby", "kind": "boundary", "nodes": [ { "id": "w-apim2", "label": "API Management", "kind": "integration", "sub": "multi-region unit", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "w-aks2", "label": "AKS Standby", "kind": "app", "sub": "min replicas", "icon": "azure_kubernetes-services" }, { "id": "w-sql2", "label": "SQL Geo-Replica", "kind": "store", "sub": "readable secondary", "icon": "azure_azure-sql" } ] }, { "title": "Execution and data standby", "kind": "boundary", "nodes": [ { "id": "w-cos2", "label": "Cosmos Read Region", "kind": "store", "sub": "failover priority 2", "icon": "azure_azure-cosmos-db" }, { "id": "w-adls2", "label": "ADLS RA-GRS", "kind": "store", "sub": "read access", "icon": "azure_datalake" }, { "id": "w-aca2", "label": "Container Apps Env", "kind": "app", "sub": "scaled to zero", "icon": "azure_container-apps-environments" } ] } ] } ], "outside": [ { "id": "w-bast", "label": "Bastion and Jump Box", "kind": "security", "sub": "no public SSH", "icon": "azure_bastions" }, { "id": "w-fw", "label": "Azure Firewall Premium", "kind": "security", "sub": "egress FQDN rules", "icon": "azure_AzureFirewall (Networking)" }, { "id": "w-mon", "label": "Azure Monitor", "kind": "platform", "sub": "workspace per region", "icon": "azure_monitor" } ], "edges": [ { "from": "w-fd", "to": "w-apim1", "label": "primary origin" }, { "from": "w-fd", "to": "w-apim2", "label": "on failure", "kind": "error", "route": "gutter" }, { "from": "w-cos1", "to": "w-cos2", "label": "replication", "kind": "async" }, { "from": "w-sql1", "to": "w-sql2", "label": "auto-failover group", "kind": "async" }, { "from": "w-aca1", "to": "w-fw", "label": "controlled egress" }, { "from": "w-adls1", "to": "w-adls2", "label": "RA-GRS", "kind": "async" } ], "note": "Executions are region-pinned for residency. Failover replays from the last checkpoint in the surviving region; it does not migrate a live run.", "meta": { "v": "1.0", "owner": "Platform Engineering", "date": "2026-08" } }, { "id": "24-cicd-and-promotion", "title": "CI/CD and Environment Promotion", "layout": "flow", "chain": true, "canvas": { "width": 1800 }, "stages": [ { "title": "Source", "nodes": [ { "id": "cd-git", "label": "GitHub Repository", "kind": "store", "sub": "platform + agents", "icon": "github" }, { "id": "cd-pr", "label": "Pull Request", "kind": "decision", "sub": "two reviewers", "icon": "github" }, { "id": "cd-iac", "label": "Bicep Modules", "kind": "app", "sub": "infrastructure", "icon": "azure_templates" } ] }, { "title": "Build", "nodes": [ { "id": "cd-img", "label": "Container Build", "kind": "app", "sub": "reproducible" }, { "id": "cd-sbom", "label": "SBOM and Scan", "kind": "security", "sub": "Defender · Trivy" }, { "id": "cd-sign", "label": "Image Signing", "kind": "security", "sub": "Notation · KV key", "icon": "azure_AzureKeyVault (Security)" } ] }, { "title": "Verify", "nodes": [ { "id": "cd-unit", "label": "Unit and Contract", "kind": "app", "sub": "schema fixtures" }, { "id": "cd-polt", "label": "Policy Simulation", "kind": "security", "sub": "Rego unit tests", "icon": "opa" }, { "id": "cd-eval", "label": "Golden Evaluation", "kind": "app", "sub": "regression gate", "icon": "azure_azure-machine-learning" }, { "id": "cd-red", "label": "Red Team Suite", "kind": "security", "sub": "injection probes" } ] }, { "title": "Gate", "nodes": [ { "id": "cd-dev", "label": "Development", "kind": "app", "sub": "auto on merge", "icon": "fa5_layer_group" }, { "id": "cd-test", "label": "Test", "kind": "app", "sub": "eval score gate", "icon": "fa5_layer_group" }, { "id": "cd-stg", "label": "Staging", "kind": "app", "sub": "load and soak", "icon": "fa5_layer_group" }, { "id": "cd-apr", "label": "Change Approval", "kind": "decision", "sub": "owner + security" } ] }, { "title": "Deploy", "nodes": [ { "id": "cd-what", "label": "What-If Plan", "kind": "app", "sub": "drift check", "icon": "azure_templates" }, { "id": "cd-bg", "label": "Blue/Green Revision", "kind": "app", "sub": "Container Apps", "icon": "azure_container-apps-environments" }, { "id": "cd-can", "label": "Canary Weight", "kind": "decision", "sub": "5 to 50 to 100" } ] }, { "title": "Watch", "nodes": [ { "id": "cd-slo", "label": "SLO Watch", "kind": "platform", "sub": "24 h window" }, { "id": "cd-qual", "label": "Online Quality", "kind": "platform", "sub": "success · escalation" }, { "id": "cd-rb", "label": "Auto Rollback", "kind": "app", "sub": "prior revision" } ] } ], "edges": [ { "from": "cd-eval", "to": "cd-test", "label": "score gate" }, { "from": "cd-slo", "to": "cd-rb", "label": "burn rate breach", "kind": "error" }, { "from": "cd-qual", "to": "cd-rb", "label": "quality regression", "kind": "error" }, { "from": "cd-can", "to": "cd-slo", "label": "each step" } ], "note": "Agent definitions travel the same pipeline as code. A prompt change is a versioned deployment, not a console edit.", "meta": { "v": "1.0", "owner": "Platform Engineering", "date": "2026-08" } }, { "id": "25-observability-and-tracing", "title": "Observability — Signal against Pipeline Stage", "layout": "grid", "canvas": { "width": 1800 }, "laneHeaderWidth": 175, "columns": [ "Emit", "Collect", "Store", "Consume", "Act" ], "rows": [ { "title": "Execution traces", "cells": [ [ { "id": "o1a", "label": "OTel GenAI spans", "kind": "app" } ], [ { "id": "o1b", "label": "OTel Collector", "kind": "integration" } ], [ { "id": "o1c", "label": "Application Insights", "kind": "store", "icon": "azure_application-insights" } ], [ { "id": "o1d", "label": "Execution graph", "kind": "platform" } ], [ { "id": "o1e", "label": "Failure analysis", "kind": "app" } ] ] }, { "title": "Metrics and SLO", "cells": [ [ { "id": "o2a", "label": "Runtime counters", "kind": "app" } ], [ { "id": "o2b", "label": "Managed Prometheus", "kind": "integration", "icon": "prometheus" } ], [ { "id": "o2c", "label": "Monitor workspace", "kind": "store", "icon": "azure_monitor" } ], [ { "id": "o2d", "label": "Managed Grafana", "kind": "platform", "icon": "azure_azure-managed-grafana" } ], [ { "id": "o2e", "label": "Page on burn rate", "kind": "security" } ] ] }, { "title": "Logs", "cells": [ [ { "id": "o3a", "label": "Structured JSON", "kind": "app" } ], [ { "id": "o3b", "label": "Container insights", "kind": "integration" } ], [ { "id": "o3c", "label": "Log Analytics", "kind": "store", "icon": "azure_log-analytics-workspaces" } ], [ { "id": "o3d", "label": "KQL incident search", "kind": "platform" } ], [ { "id": "o3e", "label": "Runbook link", "kind": "app" } ] ] }, { "title": "Token and cost", "cells": [ [ { "id": "o4a", "label": "Gateway metric", "kind": "integration", "icon": "azure_ai-gateway" } ], [ { "id": "o4b", "label": "Event Hubs", "kind": "queue", "icon": "azure_event-hubs" } ], [ { "id": "o4c", "label": "Data Explorer", "kind": "store", "icon": "azure_azure-data-explorer-clusters" } ], [ { "id": "o4d", "label": "Cost breakdown", "kind": "platform", "icon": "azure_cost-analysis" } ], [ { "id": "o4e", "label": "Budget alert", "kind": "security", "icon": "azure_cost-budgets" } ] ] }, { "title": "Guardrail decisions", "cells": [ [ { "id": "o5a", "label": "Decision event", "kind": "security" } ], [ { "id": "o5b", "label": "Event Grid", "kind": "queue", "icon": "azure_event-grid-topics" } ], [ { "id": "o5c", "label": "Audit ledger", "kind": "store", "icon": "azure_blobstorage" } ], [ { "id": "o5d", "label": "Violation report", "kind": "platform" } ], [ { "id": "o5e", "label": "Sentinel rule", "kind": "security", "icon": "azure_sentinel" } ] ] }, { "title": "Quality signals", "cells": [ [ { "id": "o6a", "label": "Judge and feedback", "kind": "app" } ], [ { "id": "o6b", "label": "Evaluation runner", "kind": "integration", "icon": "azure_azure-machine-learning" } ], [ { "id": "o6c", "label": "Eval results store", "kind": "store", "icon": "azure_azure-data-explorer-clusters" } ], [ { "id": "o6d", "label": "Agent scorecard", "kind": "platform" } ], [ { "id": "o6e", "label": "Block promotion", "kind": "decision" } ] ] } ], "note": "Execution id, workflow id, agent id and correlation id are on every row, so one identifier joins a trace to its cost and its guardrail decisions.", "meta": { "v": "1.0", "owner": "SRE", "date": "2026-08" } }, { "id": "26-cost-management", "title": "Cost Management — Meter, Attribute, Enforce, Optimise", "layout": "flow", "chain": true, "canvas": { "width": 1740 }, "stages": [ { "title": "Meter", "nodes": [ { "id": "cm-tok", "label": "Token Counting", "kind": "integration", "sub": "gateway policy" }, { "id": "cm-tool", "label": "Tool Call Units", "kind": "integration", "sub": "priced per API" }, { "id": "cm-infra", "label": "Infrastructure Cost", "kind": "platform", "sub": "Cost Management", "icon": "azure_cost-management" } ] }, { "title": "Attribute", "nodes": [ { "id": "cm-tag", "label": "Execution Tagging", "kind": "app", "sub": "tenant · project" }, { "id": "cm-roll", "label": "Rollup Job", "kind": "app", "sub": "hourly · ADX" }, { "id": "cm-alloc", "label": "Shared Cost Split", "kind": "app", "sub": "by token share" } ] }, { "title": "Enforce", "nodes": [ { "id": "cm-pre", "label": "Pre-Call Estimate", "kind": "decision", "sub": "reserve tokens" }, { "id": "cm-soft", "label": "Soft Limit", "kind": "security", "sub": "warn at 80 percent", "icon": "azure_cost-budgets" }, { "id": "cm-hard", "label": "Hard Limit", "kind": "security", "sub": "block at ceiling", "icon": "azure_cost-budgets" } ] }, { "title": "Optimise", "nodes": [ { "id": "cm-route", "label": "Cost-Aware Routing", "kind": "decision", "sub": "cheapest passing" }, { "id": "cm-cache", "label": "Cache Hit Rate", "kind": "store", "sub": "semantic + exact", "icon": "azure_redis" }, { "id": "cm-ptu", "label": "PTU Utilisation", "kind": "platform", "sub": "reserve vs burst" }, { "id": "cm-comp", "label": "Context Compression", "kind": "app", "sub": "tokens per run" } ] }, { "title": "Report", "nodes": [ { "id": "cm-dash", "label": "Cost Dashboard", "kind": "platform", "sub": "per tenant · agent", "icon": "azure_cost-analysis" }, { "id": "cm-fc", "label": "Forecast", "kind": "app", "sub": "run-rate to month end", "icon": "fa5_chart_line" }, { "id": "cm-charge", "label": "Chargeback Export", "kind": "store", "sub": "monthly · CSV", "icon": "azure_cost-management" } ] } ], "edges": [ { "from": "cm-hard", "to": "cm-route", "label": "degrade first", "kind": "error" }, { "from": "cm-fc", "to": "cm-soft", "label": "raise threshold", "route": "gutter" }, { "from": "cm-cache", "to": "cm-tok", "label": "avoided tokens", "route": "gutter" } ], "note": "A breached budget degrades the route before it blocks the run. Blocking is the last step, not the first.", "meta": { "v": "1.0", "owner": "FinOps and Architecture", "date": "2026-08" } }, { "id": "27-evaluation-quality-loop", "title": "Evaluation and Quality — The Loop That Closes", "layout": "cycle", "centre": { "label": "Agent Quality", "sub": "target 92 percent" }, "nodes": [ { "id": "q1", "label": "Capture", "kind": "store", "sub": "production traces" }, { "id": "q2", "label": "Curate", "kind": "app", "sub": "golden datasets" }, { "id": "q3", "label": "Evaluate Offline", "kind": "app", "sub": "assertions + judge", "icon": "azure_azure-machine-learning" }, { "id": "q4", "label": "Gate", "kind": "decision", "sub": "promotion score" }, { "id": "q5", "label": "Release Canary", "kind": "app", "sub": "5 percent traffic" }, { "id": "q6", "label": "Evaluate Online", "kind": "platform", "sub": "live scoring" }, { "id": "q7", "label": "Improve", "kind": "app", "sub": "prompt · model · tool" } ], "ringLabels": [ "failed and escalated runs", "labelled cases", "scores per metric", "pass or block", "canary cohort", "live signal", "new candidate version" ], "rx": 460, "ry": 230, "note": "Metrics tracked each turn: task success, groundedness, tool-call correctness, latency, tokens, cost, failure and human escalation rate.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "28-security-zones", "title": "Security Architecture — Trust Zones and What Crosses Them", "layout": "zones", "canvas": { "width": 1740 }, "zones": [ { "title": "Untrusted · internet", "kind": "trust", "nodes": [ { "id": "sz-user", "label": "Business User", "kind": "actor" }, { "id": "sz-part", "label": "Partner Webhook", "kind": "external", "sub": "HMAC signed" }, { "id": "sz-inj", "label": "Injected Content", "kind": "risk", "sub": "in retrieved page" }, { "id": "sz-att", "label": "Credential Attack", "kind": "risk", "sub": "token replay" } ] }, { "title": "Perimeter · edge", "kind": "trust", "nodes": [ { "id": "sz-ddos", "label": "DDoS Protection", "kind": "security", "sub": "network tier", "icon": "azure_ddos-protection-plans" }, { "id": "sz-waf", "label": "Front Door + WAF", "kind": "security", "sub": "OWASP ruleset", "icon": "azure_AzureFrontDoor (Networking)" }, { "id": "sz-apim", "label": "API Management", "kind": "integration", "sub": "OIDC · rate limit", "icon": "azure_AzureAPIManagement (Web)" } ] }, { "title": "Control plane · private", "kind": "trust", "nodes": [ { "id": "sz-reg", "label": "Registry API", "kind": "app", "sub": "mTLS · workload id" }, { "id": "sz-pol", "label": "Policy Service", "kind": "security", "sub": "signs bundles", "icon": "opa" }, { "id": "sz-cred", "label": "Credential Broker", "kind": "security", "sub": "no secret transit", "icon": "azure_AzureKeyVault (Security)" } ] }, { "title": "Execution plane · private, no default egress", "kind": "trust", "nodes": [ { "id": "sz-work", "label": "Runtime Workers", "kind": "app", "sub": "per-agent identity", "icon": "azure_worker-container-app" }, { "id": "sz-tool", "label": "Tool Executor", "kind": "app", "sub": "scoped token only" }, { "id": "sz-sbx", "label": "Session Sandbox", "kind": "platform", "sub": "Hyper-V · no network", "icon": "azure_azure-enclaves" }, { "id": "sz-fw", "label": "Azure Firewall", "kind": "security", "sub": "FQDN allow-list", "icon": "azure_AzureFirewall (Networking)" } ] }, { "title": "Data · restricted", "kind": "trust", "nodes": [ { "id": "sz-cos", "label": "Execution Store", "kind": "store", "sub": "CMK · private link", "icon": "azure_azure-cosmos-db" }, { "id": "sz-sql", "label": "Definition Store", "kind": "store", "sub": "CMK · row security", "icon": "azure_azure-sql" }, { "id": "sz-kv", "label": "Key Vault HSM", "kind": "security", "sub": "90 day rotation", "icon": "azure_AzureKeyVault (Security)" }, { "id": "sz-aud", "label": "Audit Ledger", "kind": "store", "sub": "WORM · 7 years", "icon": "azure_blobstorage" } ] } ], "edges": [ { "from": "sz-user", "to": "sz-waf", "label": "HTTPS 443" }, { "from": "sz-part", "to": "sz-apim", "label": "signed webhook" }, { "from": "sz-att", "to": "sz-apim", "label": "rejected at edge", "kind": "error" }, { "from": "sz-apim", "to": "sz-reg", "label": "mTLS + JWT" }, { "from": "sz-apim", "to": "sz-work", "label": "mTLS + run token" }, { "from": "sz-inj", "to": "sz-work", "label": "fenced", "kind": "error" }, { "from": "sz-cred", "to": "sz-tool", "label": "15 min token" }, { "from": "sz-work", "to": "sz-fw", "label": "egress inspected" }, { "from": "sz-work", "to": "sz-cos", "label": "private endpoint" }, { "from": "sz-pol", "to": "sz-kv", "label": "key", "route": "gutter" }, { "from": "sz-tool", "to": "sz-aud", "label": "logged", "kind": "async" } ], "note": "An agent never inherits its caller's permissions. It carries its own identity, and every zone crossing is authenticated and logged.", "meta": { "v": "1.0", "owner": "Security Architecture", "date": "2026-08" } }, { "id": "29-identity-and-delegation", "title": "Identity and Delegation — From User Sign-In to Tool Credential", "layout": "sequence", "lifelines": [ { "id": "id-user", "label": "User", "kind": "actor" }, { "id": "id-apim", "label": "API Management", "kind": "integration", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "id-entra", "label": "Microsoft Entra ID", "kind": "security", "icon": "azure_aad" }, { "id": "id-adm", "label": "Admission Service", "kind": "app" }, { "id": "id-work", "label": "Runtime Worker", "kind": "app", "icon": "azure_worker-container-app" }, { "id": "id-cred", "label": "Credential Broker", "kind": "security", "icon": "azure_AzureKeyVault (Security)" }, { "id": "id-kv", "label": "Key Vault", "kind": "security", "icon": "azure_AzureKeyVault (Security)" }, { "id": "id-tool", "label": "Downstream Tool", "kind": "external" } ], "messages": [ { "from": "id-user", "to": "id-entra", "label": "sign in, PKCE + MFA", "kind": "call" }, { "from": "id-entra", "to": "id-user", "label": "id and access token", "kind": "return" }, { "from": "id-user", "to": "id-apim", "label": "run agent, bearer", "kind": "call" }, { "from": "id-apim", "to": "id-apim", "label": "validate audience and scope", "kind": "self" }, { "from": "id-apim", "to": "id-adm", "label": "forward with caller claims", "kind": "call" }, { "from": "id-adm", "to": "id-entra", "label": "mint agent identity token", "kind": "call" }, { "from": "id-entra", "to": "id-adm", "label": "agent token, narrow scope", "kind": "return" }, { "from": "id-adm", "to": "id-work", "label": "dispatch with agent token", "kind": "async" }, { "from": "id-work", "to": "id-cred", "label": "request tool credential", "kind": "call" }, { "from": "id-cred", "to": "id-cred", "label": "check binding and approval", "kind": "self" }, { "from": "id-cred", "to": "id-kv", "label": "fetch by reference", "kind": "call" }, { "from": "id-cred", "to": "id-work", "label": "15 minute scoped token", "kind": "return" }, { "from": "id-work", "to": "id-tool", "label": "invoke with scoped token", "kind": "call" }, { "from": "id-cred", "to": "id-work", "label": "denied, not in grant", "kind": "error" } ], "note": "The user token never leaves the perimeter. An agent gets its own identity, and the intersection of user and agent grant is what it can do.", "meta": { "v": "1.0", "owner": "Security Architecture", "date": "2026-08" } }, { "id": "30-multi-tenant-isolation", "title": "Multi-Tenancy — Four Layers of Isolation", "layout": "nested", "canvas": { "width": 1760 }, "boxes": [ { "title": "Layer 1 · Identity and admission", "kind": "boundary", "dir": "row", "children": [ { "title": "Request scoping", "kind": "plain", "nodes": [ { "id": "mt-map", "label": "Tenant Mapping", "kind": "security", "sub": "Entra tenant to id", "icon": "azure_aad" }, { "id": "mt-claim", "label": "Tenant Claim", "kind": "security", "sub": "on every token", "icon": "fa5_id_badge" }, { "id": "mt-prod", "label": "APIM Product", "kind": "integration", "sub": "per tenant key", "icon": "azure_AzureAPIManagement (Web)" }, { "id": "mt-quota", "label": "Request Quota", "kind": "decision", "sub": "RPS and TPM" } ] } ] }, { "title": "Layer 2 · Runtime", "kind": "boundary", "dir": "row", "children": [ { "title": "Pooled tier", "kind": "plain", "nodes": [ { "id": "mt-pool", "label": "Shared Worker Pool", "kind": "app", "sub": "tenant-scoped run", "icon": "azure_worker-container-app" }, { "id": "mt-conc", "label": "Concurrency Cap", "kind": "decision", "sub": "per tenant" }, { "id": "mt-sess", "label": "Session per Run", "kind": "platform", "sub": "destroyed after", "icon": "azure_azure-enclaves" } ] }, { "title": "Dedicated tier", "kind": "plain", "nodes": [ { "id": "mt-ded", "label": "Dedicated Env", "kind": "app", "sub": "own Container Apps", "icon": "azure_worker-container-app" }, { "id": "mt-pe", "label": "Private Endpoints", "kind": "security", "sub": "tenant VNet", "icon": "azure_private-link" } ] } ] }, { "title": "Layer 3 · Data and secrets", "kind": "boundary", "dir": "row", "children": [ { "title": "Stores", "kind": "plain", "nodes": [ { "id": "mt-part", "label": "Cosmos Partition", "kind": "store", "sub": "tenant_id key", "icon": "azure_azure-cosmos-db" }, { "id": "mt-rls", "label": "SQL Row Security", "kind": "store", "sub": "session context", "icon": "azure_azure-sql" }, { "id": "mt-idx", "label": "Search Index", "kind": "store", "sub": "one per tenant", "icon": "azure_cognitive-search" }, { "id": "mt-cont", "label": "ADLS Container", "kind": "store", "sub": "one per tenant", "icon": "azure_datalake" } ] }, { "title": "Credentials", "kind": "plain", "nodes": [ { "id": "mt-kv", "label": "Tenant Key Vault", "kind": "security", "sub": "own CMK", "icon": "azure_AzureKeyVault (Security)" }, { "id": "mt-rot", "label": "Rotation Schedule", "kind": "security", "sub": "90 days" } ] } ] }, { "title": "Layer 4 · Telemetry, cost and audit", "kind": "boundary", "dir": "row", "children": [ { "title": "Segregation", "kind": "plain", "nodes": [ { "id": "mt-log", "label": "Log Scoping", "kind": "platform", "sub": "tenant dimension" }, { "id": "mt-adx", "label": "ADX Row Policy", "kind": "store", "sub": "tenant filter", "icon": "azure_azure-data-explorer-clusters" }, { "id": "mt-cost", "label": "Cost Attribution", "kind": "store", "sub": "tag on every call", "icon": "azure_cost-analysis" }, { "id": "mt-aud", "label": "Audit Partition", "kind": "store", "sub": "export per tenant", "icon": "azure_blobstorage" } ] } ] } ], "outside": [ { "id": "mt-noisy", "label": "Noisy Neighbour", "kind": "risk", "sub": "pooled tier only" }, { "id": "mt-cross", "label": "Cross-Tenant Read", "kind": "risk", "sub": "tested each release" } ], "edges": [ { "from": "mt-claim", "to": "mt-part", "label": "scopes every query" }, { "from": "mt-conc", "to": "mt-noisy", "label": "bounds impact", "kind": "error" }, { "from": "mt-rls", "to": "mt-cross", "label": "denies by default", "kind": "error" }, { "from": "mt-ded", "to": "mt-kv", "label": "CMK" } ], "note": "The pooled tier accepts a bounded noisy-neighbour risk in exchange for cost. The dedicated tier removes it and is priced accordingly.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } }, { "id": "31-failure-modes", "title": "Failure Modes — Signal, Classification, Response and Residual Risk", "layout": "grid", "canvas": { "width": 1800 }, "laneHeaderWidth": 180, "columns": [ "Signal", "Class", "Immediate response", "Containment", "Residual risk" ], "rows": [ { "title": "Model provider outage", "cells": [ [ { "id": "z1a", "label": "5xx and timeout", "kind": "platform" } ], [ { "id": "z1b", "label": "Provider", "kind": "decision" } ], [ { "id": "z1c", "label": "Circuit break", "kind": "integration" } ], [ { "id": "z1d", "label": "Fallback route", "kind": "app" } ], [ { "id": "z1e", "label": "Quality drift", "kind": "risk" } ] ] }, { "title": "Rate limit exhaustion", "cells": [ [ { "id": "z2a", "label": "429 burst", "kind": "platform" } ], [ { "id": "z2b", "label": "Transient", "kind": "decision" } ], [ { "id": "z2c", "label": "Backoff + jitter", "kind": "queue" } ], [ { "id": "z2d", "label": "PTU spillover", "kind": "app" } ], [ { "id": "z2e", "label": "Latency spike", "kind": "risk" } ] ] }, { "title": "Tool unavailable", "cells": [ [ { "id": "z3a", "label": "Health probe red", "kind": "platform" } ], [ { "id": "z3b", "label": "Tool", "kind": "decision" } ], [ { "id": "z3c", "label": "Retry then skip", "kind": "app" } ], [ { "id": "z3d", "label": "Degraded answer", "kind": "app" } ], [ { "id": "z3e", "label": "Incomplete result", "kind": "risk" } ] ] }, { "title": "Agent loop divergence", "cells": [ [ { "id": "z4a", "label": "Step count rising", "kind": "platform" } ], [ { "id": "z4b", "label": "Agent or model", "kind": "decision" } ], [ { "id": "z4c", "label": "Depth and token cap", "kind": "security" } ], [ { "id": "z4d", "label": "Terminate run", "kind": "app" } ], [ { "id": "z4e", "label": "Wasted spend", "kind": "risk" } ] ] }, { "title": "Poison message", "cells": [ [ { "id": "z5a", "label": "Delivery count", "kind": "platform" } ], [ { "id": "z5b", "label": "Permanent", "kind": "decision" } ], [ { "id": "z5c", "label": "Dead letter", "kind": "queue", "icon": "azure_AzureServiceBusQueue (Integration)" } ], [ { "id": "z5d", "label": "Triage and replay", "kind": "app" } ], [ { "id": "z5e", "label": "Manual backlog", "kind": "risk" } ] ] }, { "title": "Policy or auth denial", "cells": [ [ { "id": "z6a", "label": "403 on tool call", "kind": "security" } ], [ { "id": "z6b", "label": "Policy or auth", "kind": "decision" } ], [ { "id": "z6c", "label": "Fail closed", "kind": "security" } ], [ { "id": "z6d", "label": "Escalate to human", "kind": "actor" } ], [ { "id": "z6e", "label": "Workflow stalls", "kind": "risk" } ] ] }, { "title": "Region loss", "cells": [ [ { "id": "z7a", "label": "Probe failure", "kind": "platform" } ], [ { "id": "z7b", "label": "Infrastructure", "kind": "decision" } ], [ { "id": "z7c", "label": "Front Door reroute", "kind": "integration" } ], [ { "id": "z7d", "label": "Replay checkpoints", "kind": "app" } ], [ { "id": "z7e", "label": "RPO 5 min gap", "kind": "risk" } ] ] }, { "title": "Control plane outage", "cells": [ [ { "id": "z8a", "label": "Registry 5xx", "kind": "platform" } ], [ { "id": "z8b", "label": "Infrastructure", "kind": "decision" } ], [ { "id": "z8c", "label": "Admissions stop", "kind": "security" } ], [ { "id": "z8d", "label": "Cached bundles run", "kind": "app" } ], [ { "id": "z8e", "label": "No new versions", "kind": "risk" } ] ] } ], "note": "Every row ends in a residual risk that is accepted, not eliminated. Those risks are the agenda for the operational review.", "meta": { "v": "1.0", "owner": "SRE and Architecture", "date": "2026-08" } }, { "id": "32-delivery-roadmap", "title": "Delivery Roadmap — Foundation, MVP and Two Phases", "layout": "flow", "chain": true, "canvas": { "width": 1780 }, "stages": [ { "title": "Foundation · 6 weeks", "nodes": [ { "id": "rm-lz", "label": "Landing Zone", "kind": "platform", "sub": "subscription · Bicep" }, { "id": "rm-net", "label": "Network and Private Link", "kind": "security", "sub": "no public data path", "icon": "azure_private-link" }, { "id": "rm-idn", "label": "Entra and RBAC", "kind": "security", "sub": "roles · workload id", "icon": "azure_aad" }, { "id": "rm-cicd", "label": "CI/CD Skeleton", "kind": "app", "sub": "build · sign · deploy" } ] }, { "title": "MVP core · 12 weeks", "nodes": [ { "id": "rm-areg", "label": "Agent Registry", "kind": "app", "sub": "versioned specs" }, { "id": "rm-rt", "label": "Agent Runtime", "kind": "app", "sub": "single agent" }, { "id": "rm-wf", "label": "Workflow Engine", "kind": "app", "sub": "sequential + branch" }, { "id": "rm-gw", "label": "Model Gateway", "kind": "integration", "sub": "one provider pool", "icon": "azure_ai-gateway" }, { "id": "rm-tool", "label": "Tool Registry", "kind": "app", "sub": "REST + MCP", "icon": "azure_api-center" } ] }, { "title": "MVP controls · with core", "nodes": [ { "id": "rm-dur", "label": "Durable State", "kind": "store", "sub": "checkpoint · resume" }, { "id": "rm-grd", "label": "Guardrails", "kind": "security", "sub": "in · out · tool", "icon": "azure_content-safety" }, { "id": "rm-trc", "label": "Execution Tracing", "kind": "platform", "sub": "OTel GenAI" }, { "id": "rm-cost", "label": "Cost Tracking", "kind": "store", "sub": "token accounting", "icon": "azure_cost-analysis" }, { "id": "rm-api", "label": "API, SDK, Console", "kind": "app", "sub": "the whole surface", "icon": "python" } ] }, { "title": "Phase 2 · 2 quarters", "nodes": [ { "id": "rm-multi", "label": "Multi-Agent Patterns", "kind": "app", "sub": "six patterns" }, { "id": "rm-hitl", "label": "Human-in-the-Loop", "kind": "app", "sub": "approval queues" }, { "id": "rm-mem", "label": "Advanced Memory", "kind": "store", "sub": "episodic · semantic", "icon": "fa5_memory" }, { "id": "rm-route", "label": "Routing and Failover", "kind": "integration", "sub": "policy driven" }, { "id": "rm-eval", "label": "Evaluation Framework", "kind": "app", "sub": "offline + online", "icon": "azure_azure-machine-learning" }, { "id": "rm-ten", "label": "Multi-Tenancy", "kind": "security", "sub": "four layers", "icon": "azure_multi-tenancy" } ] }, { "title": "Phase 3 · beyond", "nodes": [ { "id": "rm-disc", "label": "Agent Discovery", "kind": "app", "sub": "dynamic delegation" }, { "id": "rm-plan", "label": "Autonomous Planning", "kind": "app", "sub": "replanning" }, { "id": "rm-region", "label": "Multi-Region Execution", "kind": "platform", "sub": "residency routing" }, { "id": "rm-pac", "label": "Policy as Code", "kind": "security", "sub": "tenant authored", "icon": "opa" }, { "id": "rm-mkt", "label": "Agent Marketplace", "kind": "app", "sub": "internal catalogue" } ] } ], "edges": [ { "from": "rm-dur", "to": "rm-hitl", "label": "prerequisite", "route": "gutter" }, { "from": "rm-trc", "to": "rm-eval", "label": "prerequisite", "route": "gutter" }, { "from": "rm-ten", "to": "rm-region", "label": "prerequisite", "route": "gutter" } ], "note": "Durable state, tracing and tenancy are prerequisites, not enhancements. Deferring them makes every later phase more expensive.", "meta": { "v": "1.0", "owner": "Data & AI Global Practice", "date": "2026-08" } } ]