Microsoft AI-500: Architect multi-agent solutions Practice Questions

Option A is correct because Azure API Management (APIM) serves as the enterprise API gateway: it enforces OAuth 2.0 token validation, validates incoming/outgoing request schemas, applies per-agent rate limits (TPM/RPM), and centralizes logging before forwarding calls to downstream backend services. Option B is incorrect because Event Hubs is an event-ingestion streaming engine, not an API gateway. Option C is incorrect because Service Bus Relay securely exposes on-premises services without schema enforcement. Option D is incorrect because a WAF in detection mode only logs suspected web attack payloads without enforcing API token policies or schema validation.

Option B is correct because the Agent-to-Agent (A2A) protocol is the industry standard supported by Microsoft for multi-agent interoperability; agents publish standardized 'Agent Cards' detailing their capabilities, operational schemas, and endpoints so orchestrators can discover and invoke them across heterogeneous platforms. Option A is incorrect because MCP is specifically designed for agent-to-tool and agent-to-data-resource integration (client-to-server), not agent-to-agent capability discovery and coordination. Option C is incorrect because OpenAPI specifies static HTTP APIs, not agent cognitive capabilities or dynamic handoff contracts. Option D is incorrect because WebSub is an asynchronous pub-sub webhook protocol for web feeds, not agent capability discovery.

Option A is correct because the Blackboard architecture features independent, heterogeneous specialized agents working collaboratively and asynchronously by posting partial findings and hypotheses to a shared common data store (the blackboard) which other agents inspect and refine. Option B is incorrect because sequential pipelines pass data linearly from agent to agent. Option C is incorrect because master-worker relies on a single master directly assigning subtasks to workers. Option D is incorrect because a broadcast bus streams messages to all listeners without managing a persistent shared problem-solving blackboard.

Option B is correct because low-latency in-memory stores like Azure Cache for Redis handle high-frequency active session working memory, while durable databases like Azure Cosmos DB with partition keys based on tenantId provide secure, scalable, tenant-isolated long-term persistence. Option A is incorrect because dumping all history and policies into the system prompt exhausts token limits and increases costs drastically. Option C is incorrect because public Blob Storage violates Zero Trust and data confidentiality. Option D is incorrect because browser localStorage exposes sensitive tenant state to client-side tampering and lacks backend durability.

Option A is correct because the W3C Distributed Tracing specification defines standard traceparent (Trace ID and Span ID) and tracestate headers that must be propagated across all inter-service and inter-agent calls, enabling Azure Monitor and Azure AI Foundry to reconstruct the full distributed flame graph. Option B, C, and D are incorrect because they provide authentication or encryption mechanisms, not distributed trace correlation.