Environment & Operations › Activity inventory
Activity inventory
Activity inventory
Identity, access & tenancy
- P Provision and rotate agent service identities, each agent authenticates as itself
- P Per-tenant isolation: separate stores, keys, network boundaries; automated cross-tenant read-path verification
- P SSO / OAuth identity-provider integration, session issuance, token refresh
- H Permission-scope enforcement at the MCP layer (mechanism only, scope content comes from the plane)
AWSAgentCore Identity · IAM Identity Center · Cognito · KMS
Updates here: IdP federation per plane, one service identity per agent, per-tenant CMKs, the automated cross-tenant isolation test.
Updates here: IdP federation per plane, one service identity per agent, per-tenant CMKs, the automated cross-tenant isolation test.
AZUREMicrosoft Entra ID · managed identities · Key Vault
Updates here: tenant model on Entra, one service principal / managed identity per agent, per-tenant keys, the automated isolation test.
Updates here: tenant model on Entra, one service principal / managed identity per agent, per-tenant keys, the automated isolation test.
GCPCloud IAM · per-agent service accounts · Workload Identity
Updates here: tenant model across projects, one service account per agent, per-tenant CMEK, the automated isolation test.
Updates here: tenant model across projects, one service account per agent, per-tenant CMEK, the automated isolation test.
Runtime & infrastructure operations
- P Cloud provisioning, containers, queues, caching, secrets, certificates, backup and restore
- P CI/CD pipelines, environment promotion, rollback
- P Repo scaffolding: standard project skeleton, new engagements start from template, never blank
- P Uptime, capacity and incident management for hosted MCP servers and the orchestrator
- P Disaster recovery: versioned backup and rehearsed restore of tenant stores and configuration
AWSAWS Organizations · VPC · CodePipeline · Terraform/CDK
Updates here: account structure & SCPs, network baseline, SSRA skeleton stamping pipeline, runbooks.
Updates here: account structure & SCPs, network baseline, SSRA skeleton stamping pipeline, runbooks.
AZUREAzure subscriptions & management groups · Bicep/Terraform · Azure DevOps
Updates here: subscription structure, network baseline, SSRA skeleton stamping pipeline, runbooks.
Updates here: subscription structure, network baseline, SSRA skeleton stamping pipeline, runbooks.
GCPGCP organization / folders / projects · Terraform · Cloud Build
Updates here: project structure, network baseline, SSRA skeleton stamping pipeline, runbooks.
Updates here: project structure, network baseline, SSRA skeleton stamping pipeline, runbooks.
MCP machinery: the north gateway and the south tool servers
- H North-facing Platform MCP Gateway: per-tenant endpoints, tenant auth, host allow-list enforcement, identity passthrough, rate limits and per-host quotas
- H Server hosting, health checks, pooling, retry/backoff, rate limits, timeouts
- H Connector authentication handshakes to source systems (mechanics only)
- H Tool registry and schema validation on every deploy
- H Uniform pagination, truncation and response-size defaults
AWSAgentCore Runtime (north MCP serving) · AgentCore Gateway (south tool servers)
Updates here: per-tenant north endpoint template + host allow-list enforcement; south server registrations, tool schemas, permission scopes.
Updates here: per-tenant north endpoint template + host allow-list enforcement; south server registrations, tool schemas, permission scopes.
AZUREAzure API Management (north per-tenant endpoints) · AI Foundry Agent Service tool calling + APIM (south)
Updates here: per-tenant north endpoint policies + host allow-list; south tool registrations, schemas, scopes.
Updates here: per-tenant north endpoint policies + host allow-list; south tool registrations, schemas, scopes.
GCPApigee (north per-tenant endpoints) · Vertex AI Agent Engine tools + Apigee (south)
Updates here: per-tenant north endpoint proxies + host allow-list; south tool registrations, schemas, scopes.
Updates here: per-tenant north endpoint proxies + host allow-list; south tool registrations, schemas, scopes.
Intelligence-layer machinery (operates L4)
- H Model router: vendor credentials, version pinning, fallback chains, A/B routing
- H The single model-boundary abstraction, a vendor or version swap touches only this
- H Embedding and reranker service operation
- H Model deprecation tracking; regression evals before any version migration
AWSAmazon Bedrock · inference profiles
Updates here: model router config, version pins & fallback chains, regression-eval gate before any migration, cost ceilings from the plane.
Updates here: model router config, version pins & fallback chains, regression-eval gate before any migration, cost ceilings from the plane.
AZUREAzure AI Foundry model catalog (Claude available via Foundry)
Updates here: router config, deployment pins & fallbacks, regression-eval gate, cost ceilings from the plane.
Updates here: router config, deployment pins & fallbacks, regression-eval gate, cost ceilings from the plane.
GCPVertex AI Model Garden (Claude available on Vertex)
Updates here: router config, version pins & fallbacks, regression-eval gate, cost ceilings from the plane.
Updates here: router config, version pins & fallbacks, regression-eval gate, cost ceilings from the plane.
Observability, tracing & metering
- H Content-blind production tracing of every run, structure recorded, deal data never
- P Log aggregation, retention enforcement, append-only immutable audit-trail storage, an entry is never updated in place
- H Deliverable lineage: every generated output logged with its template version and status, artifacts as traceable as the values inside them
- H Cost and token metering per run, per agent, per tenant, against each run’s declared budget
- H Latency tracking per tool call and agent chain; regression alerting
- H Gate-metrics collection, the numbers; interpretation is solution-layer work
AWSAgentCore Observability · CloudWatch · OTEL · S3 Object Lock
Updates here: content-blind trace configuration (verified), cost-vs-budget dashboards, latency alerting, immutable audit storage.
Updates here: content-blind trace configuration (verified), cost-vs-budget dashboards, latency alerting, immutable audit storage.
AZUREAzure Monitor · Application Insights · OTEL · immutable Blob storage
Updates here: content-blind trace config, cost-vs-budget dashboards, latency alerting, immutable audit store.
Updates here: content-blind trace config, cost-vs-budget dashboards, latency alerting, immutable audit store.
GCPCloud Trace / Logging / Monitoring · OTEL · GCS Bucket Lock
Updates here: content-blind trace config, cost-vs-budget dashboards, latency alerting, immutable audit store.
Updates here: content-blind trace config, cost-vs-budget dashboards, latency alerting, immutable audit store.
Evaluation infrastructure: the runner lives here, the rubrics live in Knowledge
- H Eval runner: scheduled and pre-deploy suite execution, score storage, trend dashboards
- H LLM-as-judge execution machinery
- H Tool-layer eval harness: accuracy, call count, error rate per MCP server
- H Clean-instance comparison tooling: run both variants, diff outputs and token spend
AWSAgentCore Evaluations · custom eval runners (Lambda/Step Functions)
Updates here: golden-exemplar suite wiring, LLM-as-judge templates, tool-layer eval harness, clean-instance A/B tooling.
Updates here: golden-exemplar suite wiring, LLM-as-judge templates, tool-layer eval harness, clean-instance A/B tooling.
AZUREAI Foundry Evaluations · custom runners (Azure Functions)
Updates here: golden-exemplar suites, judge templates, tool-layer evals, clean-instance A/B.
Updates here: golden-exemplar suites, judge templates, tool-layer evals, clean-instance A/B.
GCPVertex AI Evaluation Service · custom runners (Cloud Functions)
Updates here: golden-exemplar suites, judge templates, tool-layer evals, clean-instance A/B.
Updates here: golden-exemplar suites, judge templates, tool-layer evals, clean-instance A/B.
Guardrail enforcement machinery
- H Hooks engine: deterministic pre/post checks, zero exceptions
- H Permission-scope enforcement and violation logging
- H Gate-workflow mechanics: transitions blocked until approval recorded; escalation timers
- H PII detection and redaction as a service
AWSAgentCore Policy (Cedar) · Bedrock Guardrails · Comprehend PII · Runtime hooks
Updates here: Cedar encodings of the four locks, gate-blocking mechanics, PII detection/redaction pipeline.
Updates here: Cedar encodings of the four locks, gate-blocking mechanics, PII detection/redaction pipeline.
AZUREAzure AI Content Safety · Microsoft Purview · policy hooks
Updates here: lock encodings, gate-blocking mechanics, PII detection/redaction pipeline.
Updates here: lock encodings, gate-blocking mechanics, PII detection/redaction pipeline.
GCPModel Armor · Sensitive Data Protection (DLP) · policy hooks
Updates here: lock encodings, gate-blocking mechanics, PII detection/redaction pipeline.
Updates here: lock encodings, gate-blocking mechanics, PII detection/redaction pipeline.
Session & context machinery
- H Context-window management: compaction execution, session lifecycle
- H Memory stores: short-term working memory, long-term store operation
- H Context-pack delivery: assembly, caching, size enforcement (contents come from Knowledge)
- H Loop enforcement: kill any run exceeding its declared token budget or iteration bound; raise an incident
- H Long-horizon state: agents summarize completed phases into external memory and resume from checkpoints instead of restarting when a run is interrupted
AWSAgentCore Memory · Runtime · ElastiCache
Updates here: short/long-term memory stores, compaction policy, context-pack assembly cache, run-contract budget-kill enforcement.
Updates here: short/long-term memory stores, compaction policy, context-pack assembly cache, run-contract budget-kill enforcement.
AZUREAI Foundry threads & memory · Azure Cache for Redis
Updates here: memory stores, compaction policy, context-pack cache, budget-kill enforcement.
Updates here: memory stores, compaction policy, context-pack cache, budget-kill enforcement.
GCPAgent Engine sessions & memory · Memorystore
Updates here: memory stores, compaction policy, context-pack cache, budget-kill enforcement.
Updates here: memory stores, compaction policy, context-pack cache, budget-kill enforcement.
Data-layer machinery (operates L5)
- P Ingestion mechanics: OCR, format conversion, chunking, embedding, tenant-store indexing
- P Sync schedulers, change detection, dead-letter handling
- P Storage lifecycle execution: retention, legal hold, secure deletion (policies come from the plane)
- P Scoped index lifecycle: per-engagement retrieval indexes created at case open, destroyed at case close, never outliving their source entitlements
- P Restricted-store enforcement: walled sources excluded from shared indexes and pipelines; entitlement-checked retrieval only
AWSS3 (per-tenant) · Bedrock Knowledge Bases · AWS Glue · OpenSearch Serverless · S3 Vectors · Aurora PostgreSQL · Neo4j Aura (or Neptune) · Kinesis · Step Functions
Updates here: Bedrock Knowledge Bases as the managed ingestion path (chunking, embedding, indexing, sync over S3 data sources); Glue ETL jobs & crawlers with plane field-mappings where custom processing is required; per-tenant store provisioning templates; ontology release load into the graph (Bolt loader); streaming ingestion; SQS dead-letter queues. Residency caution: Knowledge Bases are region-scoped, and S3 data resides in one region even though the bucket namespace is global; both must match the C6 data_residency value.
Updates here: Bedrock Knowledge Bases as the managed ingestion path (chunking, embedding, indexing, sync over S3 data sources); Glue ETL jobs & crawlers with plane field-mappings where custom processing is required; per-tenant store provisioning templates; ontology release load into the graph (Bolt loader); streaming ingestion; SQS dead-letter queues. Residency caution: Knowledge Bases are region-scoped, and S3 data resides in one region even though the bucket namespace is global; both must match the C6 data_residency value.
AZUREBlob Storage · Azure Data Factory · Azure AI Search · Azure SQL / Microsoft Fabric · Neo4j Aura · Event Hubs / Service Bus
Updates here: ADF pipelines with plane field-mappings, per-tenant provisioning templates, ontology load (Bolt loader), streaming, dead-letter queues.
Updates here: ADF pipelines with plane field-mappings, per-tenant provisioning templates, ontology load (Bolt loader), streaming, dead-letter queues.
GCPGCS · Dataflow / Data Fusion · Vertex AI Vector Search · BigQuery / AlloyDB · Neo4j Aura · Pub/Sub
Updates here: pipelines with plane field-mappings, per-tenant provisioning templates, ontology load (Bolt loader), streaming, dead-letter topics.
Updates here: pipelines with plane field-mappings, per-tenant provisioning templates, ontology load (Bolt loader), streaming, dead-letter topics.
Version & release mechanics
- P Semver tooling, changelog generation, referential-integrity validation on every commit
- P Artifact regeneration pipelines, generated views, never hand-edited, enforced by CI
- P Dependency and vulnerability scanning
- P Rainbow deployments: traffic shifts gradually from the old version to the new so running agents are never disrupted mid-flight
AWSCodePipeline / GitHub Actions · Terraform
Updates here: CI gates (business-blindness linter, referential integrity, schema validation, eval regression), artifact regeneration pipelines.
Updates here: CI gates (business-blindness linter, referential integrity, schema validation, eval regression), artifact regeneration pipelines.
AZUREAzure DevOps / GitHub Actions · Terraform
Updates here: the five CI gates, artifact regeneration pipelines.
Updates here: the five CI gates, artifact regeneration pipelines.
GCPCloud Build / GitHub Actions · Terraform
Updates here: the five CI gates, artifact regeneration pipelines.
Updates here: the five CI gates, artifact regeneration pipelines.