A reference implementation of an operations assistant on Azure that enforces human approval before executing mutations. The assistant retrieves runbooks via Azure AI Search using hybrid lexical+vector search with semantic reranking, calls Azure OpenAI for reasoning, and separates read-only tool calls from write operations. When the model proposes an action like opening an incident, it records a pending action and returns a preview — execution only happens after the authenticated user explicitly confirms. Key engineering decisions covered include: identity-scoped session isolation using Container Apps Entra ID auth headers, managed identity for keyless Azure service access, idempotent confirmation handling with in-process locking, separate telemetry vs. audit streams, and honest acknowledgment of what the single-replica in-memory design cannot yet handle at scale.
