Azure expert

Azure Expert

Azure is policy-first: Entra ID and Azure Policy are the guardrails; the Foundry is the agent surface. If it is not enforced by policy, it is a wish.

This skill enforces the discipline that makes Azure workloads production-safe: identity, policy, data boundaries, cost controls, and residency. It is not an Azure feature tour — it is a checklist of the things that cause incidents and compliance failures when skipped.

When to use

• Designing any Azure infrastructure (new or modified)
• Before deploying agents or LLM workloads to Azure (AI Foundry, Azure OpenAI, Azure AI Agent Service)
• When reviewing a Bicep/Terraform plan or a landing-zone design
• When a system spans subscriptions, touches regulated data, or crosses geographies

Procedure

1. Identity and access — verify least-privilege for every identity:

   • Managed identities over service principals with secrets; no client secrets in code or config.
   • Entra ID RBAC scoped to the specific function; PIM for standing privileged access.
   • Conditional Access policies for human principals on production subscriptions.

2. Policy and landing zones — confirm governance is mechanical:

   • Azure Policy assignments enforce allowed locations, required encryption, and denied public endpoints.
   • Workload sits inside a Cloud Adoption Framework landing zone (or an explicit, owned deviation).

3. Data boundaries — for every data store:

   • Classification recorded (Purview where in scope); CMK where required.
   • Private Endpoints on PaaS data services; public network access disabled by default.
   • Cross-tenant or cross-subscription sharing explicit and documented.

4. Data residency — for each resource:

   • Allowed-locations policy constrains deployment geography (e.g. EU Data Boundary where required).
   • For Azure OpenAI / Foundry calls: regional deployments, not global, where residency matters.

5. Cost controls — for every LLM, compute, or storage resource:

   • Cost Management budgets with alerts at 50%, 75%, 90%, 100%.
   • Foundry model quotas and rate limits set; autoscale maximums bounded.

6. Network and egress — confirm:

   • Private Link over public endpoints; NSGs default-deny.
   • Egress costed for cross-region and internet-bound traffic.

7. Observability — confirm:

   • Azure Monitor / Log Analytics dashboards and alerts on error rate, latency, and cost.
   • Activity logs routed centrally for retention and audit — evidence before go-live.

8. Run the Adversarial Gate — common Azure failure modes: service principals with secrets, public PaaS endpoints, missing allowed-locations policy, uncapped Foundry spend, standing owner access without PIM.

Official sources — validate before you assert

• Documentation: learn.microsoft.com/azure · Well-Architected: learn.microsoft.com/azure/well-architected
• Live docs via MCP (verify currency before pinning): the official Azure MCP Server (github.com/microsoft/azure-mcp) and the Microsoft Learn docs MCP endpoint — fetch current docs instead of relying on memory.
• GitHub, foundations: github.com/Azure/Enterprise-Scale (CAF landing zones) · github.com/Azure/bicep · github.com/Azure-Samples
• GitHub, agent examples: Azure AI Agent Service samples under github.com/Azure-Samples · AutoGen (github.com/microsoft/autogen) · Semantic Kernel (github.com/microsoft/semantic-kernel)
• Rule: every service/API claim cites an official doc. Quotas, prices, and model names are dated facts — stale until re-verified against the source.

Outputs

• Azure guardrail checklist (pass/fail per item)
• Identity matrix: principal | role | scope | justification
• Data classification and boundary map
• Budget alert confirmation
• Open findings for human review

Guardrails

• No client secrets in code or config. Managed identity or nothing.
• Policy, not promises. If a guardrail is not an Azure Policy assignment, it does not exist.
• Residency is a constraint, not a preference. Enforce with allowed-locations policy.
• Budget alerts are not optional. An unmonitored LLM workload will produce a surprise invoice.