Cloud platform overlays
Select a platform to see every environment activity mapped to that platform's services. The portable shelf below the maps lists the cross-cloud choices that shrink the swap surface no matter which platform is selected.
AWS component map: where the updates are made
AgentCore Runtime + Harness
Deploy orchestrator & 14 agents in isolated microVMs; serve them northward over MCP (per-tenant endpoints); enforce run contracts, over-budget loops killed.
AgentCore Gateway
Register the 10 logical south tool servers (~4 physical); tool schemas validated in CI; permission scopes defined once per server, values from the plane.
AgentCore Identity
Federate the client IdP (Entra/Cognito/Okta per plane); mint one service identity per agent; OAuth grants to source systems.
AgentCore Memory
Short-term session + long-term stores; compaction policy config (what survives summarization).
AgentCore Policy (Cedar)
Encode the enforcement half of the four guardrail locks; gate transitions blocked until approval recorded; scope-violation logging.
AgentCore Observability
OTEL traces to CloudWatch; content-blind configuration verified; dashboards: cost-vs-budget, latency, gate metrics per agent per tenant.
AgentCore Evaluations
Wire golden-exemplar suites and LLM-as-judge pipelines; pre-deploy eval gates on agents and tools.
AgentCore Registry
Generated from the solution repo, never hand-edited; registry-only credential issuance makes unregistered agents inert.
AgentCore Code Interpreter
Sandboxed execution so agents verify computations instead of asserting them.
Amazon Bedrock (models)
Model access + inference profiles; router pins, fallbacks and A/B config, the portability boundary lives in this config.
S3 (per-tenant)
Tenant document stores: bucket/prefix isolation, per-tenant KMS, lifecycle & retention rules executed from plane values. Data resides in one region even though the namespace is global; residency follows the C6 value.
Bedrock Knowledge Bases
The managed ingestion path: chunking, embedding, indexing and sync over per-tenant S3 data sources; external sources such as Google Drive connect through gateway-managed OAuth. Default for POCs and standard engagements; manual or periodic sync is POC-only, production sync is change-detection driven with dead-letter handling.
OpenSearch Serverless
Per-tenant vector collections for low-latency, high-throughput retrieval; index templates; retrieval API backing.
S3 Vectors
Storage-first vector option for massive, cost-optimized collections at subsecond query; Bedrock Knowledge Bases can back onto either this or OpenSearch. Choose by latency and query volume, not by raw scale.
Aurora PostgreSQL
SQL/warehouse service; per-tenant schemas; taxonomy reference tables; optional pgvector for smaller tenants.
Neo4j Aura (or Neptune)
The Knowledge graph store: load each pinned ontology release via the Bolt loader; read-only role for agents; version label on every node.
AWS Glue
ETL catalog & jobs: chunking/embedding pipelines; schema mappings materialized from plane field-maps; crawlers over the source estate.
Kinesis / MSK
Real-time streaming ingestion where sources emit events; sync schedules from the plane via EventBridge.
Step Functions + Lambda
Deterministic workflow engine (the 28 workflow-disposition capabilities) and ingestion orchestration; SQS dead-letter queues.
IAM Identity Center + Cognito
Human SSO; tenant authentication for north-gateway endpoints; host allow-list checks.
KMS + Secrets Manager
Per-tenant customer-managed keys; source credentials from the plane, stored with rotation schedules, never in code.
CloudWatch + OTEL
Log aggregation, retention enforcement, immutable audit trail (S3 Object Lock); the operational loop's telemetry source.
CodePipeline / Terraform
CI/CD with the five rogue-development gates; SSRA skeleton stamping; environment promotion dev→staging→prod.
Azure component map: where the updates are made
Azure AI Foundry Agent Service
Agent runtime: orchestrator & agents, threads, tool calling; serve northward over MCP behind APIM per-tenant endpoints; run-contract enforcement in the loop.
Azure API Management
Per-tenant north endpoints with tenant auth + host allow-list policies; can also front the south tool servers.
Microsoft Entra ID
Human SSO and agent service principals / managed identities; OAuth grants to source systems. Often ALREADY the client's IdP, a real Azure advantage.
AI Foundry model catalog
Model deployments + router pins & fallbacks; Claude available via Foundry, the portability boundary lives in this config.
Azure AI Content Safety + Purview
Guardrail enforcement half: content filters, PII detection, sensitivity labels, compliance posture.
Azure Monitor + App Insights
OTEL tracing, content-blind config, dashboards; immutable audit via Blob immutability policies.
AI Foundry Evaluations
Eval runner: golden-exemplar suites, judge pipelines, pre-deploy gates.
Blob Storage + Azure AI Search
Per-tenant document stores and vector/hybrid search indexes.
Azure SQL / Microsoft Fabric
Warehouse & SQL service; per-tenant schemas; Fabric if the client is Microsoft-centric end-to-end.
Neo4j Aura (external)
Same graph store on every cloud, the Knowledge layer never moves when the cloud does.
Azure Data Factory
ETL pipelines; schema mappings from plane field-maps; source-estate connectors.
Event Hubs / Service Bus
Streaming ingestion, queues and dead-lettering; schedules via Logic Apps.
Logic Apps / Durable Functions
Deterministic workflow engine and ingestion orchestration.
Key Vault
Per-tenant keys and source credentials with rotation, never in code.
Azure DevOps / GitHub Actions + Terraform
CI/CD with the five rogue-development gates; SSRA skeleton stamping.
GCP component map: where the updates are made
Vertex AI Agent Engine
Agent runtime (Agent Builder / ADK): orchestrator & agents, sessions, tool calling; MCP and A2A support; run-contract enforcement in the loop.
Apigee
Per-tenant north endpoint proxies with tenant auth + host allow-list; can front the south tool servers.
Cloud IAM + Workload Identity
Human SSO federation and per-agent service accounts; OAuth grants to source systems.
Vertex AI Model Garden
Model deployments + router pins & fallbacks; Claude available on Vertex, the portability boundary lives in this config.
Model Armor + Sensitive Data Protection
Guardrail enforcement half: prompt/response screening, PII inspection and de-identification.
Cloud Trace / Logging / Monitoring
OTEL tracing, content-blind config, dashboards; immutable audit via GCS Bucket Lock.
Vertex AI Evaluation Service
Eval runner: golden-exemplar suites, judge pipelines, pre-deploy gates.
GCS + Vertex AI Vector Search
Per-tenant document stores and vector indexes; Firestore for lightweight document state.
BigQuery / AlloyDB
Warehouse (BigQuery is GCP's strongest data card) and transactional SQL; per-tenant datasets.
Neo4j Aura (external)
Same graph store on every cloud, the Knowledge layer never moves when the cloud does.
Dataflow / Data Fusion
ETL and streaming pipelines; schema mappings from plane field-maps.
Pub/Sub
Event streaming, schedules and dead-letter topics.
Cloud Workflows / Functions
Deterministic workflow engine and ingestion orchestration.
Secret Manager + Cloud KMS
Per-tenant CMEK and source credentials with rotation, never in code.
Cloud Build / GitHub Actions + Terraform
CI/CD with the five rogue-development gates; SSRA skeleton stamping.
Portable by design: the anti-lock-in shelf, common across all three
| Tool / product | Role | Why it keeps you portable |
|---|---|---|
| Neo4j Aura | Knowledge graph | Identical on all three clouds, the ontology never migrates |
| Terraform | IaC | One IaC language across AWS/Azure/GCP, the swap surface |
| GitHub Actions | CI/CD | Cloud-neutral pipelines; the five CI gates written once |
| OpenTelemetry | Tracing | One instrumentation standard; only the sink changes per cloud |
| Airbyte / Fivetran | Connector library | Cross-cloud connector ecosystems, an alternative or complement to native ETL |
| Kafka / Confluent | Streaming | If streaming must be portable, prefer over native buses |
| Qdrant / Pinecone / Weaviate | Vector store | Cloud-neutral vector options if OpenSearch/AI Search/Vertex lock-in is a concern |
| Temporal | Workflow engine | Portable deterministic orchestration alternative to Step Functions/Logic Apps/Workflows |
| Okta | Identity | Cloud-neutral IdP when the client is not already Entra/Google |
| LangFuse | LLM observability | Open-source run tracing/evals overlay, portable across model vendors |
The portability rule: anything above the Environment layer never learns which cloud it runs on. These cross-cloud choices shrink the swap surface further, the cloud selector above changes the bottom of the stack, never the platform.
The harness runtime itself can also be rented rather than built: Anthropic Managed Agents, AWS AgentCore Runtime, Azure AI Foundry Agent Service and Google Vertex AI Agent Engine each operate the agent loop, sandbox and session state as a managed service. Two cautions if we take that path: renting the harness from a model vendor moves the lock-in question from the Intelligence layer to the Environment layer, so the harness choice must never constrain the model choice; and the Managed Agents beta currently carries no Zero Data Retention or HIPAA coverage, which the data-sensitivity boundary must respect. The buy-versus-build call is an open item on the Decision Register.