Design enterprise-scale agent communication with A2A in Azure

Why it matters

Independently deployed agents need discovery, isolation, and shared state that do not depend on hardcoded endpoints or on what a remote agent claims about itself.

What to practise

Complete a local A2A-compatible server that publishes /.well-known/agent-card.json and accepts JSON-RPC messages that reach a Foundry agent. Register agent cards in an Azure Cosmos DB container partitioned by tenant, with heartbeats and a TTL. Discover the lowest-load healthy agent for a capability, update shared task state with ETags and a bounded 412 retry, and write conflict decisions to a durable audit container. Watch that the tenant comes from trusted configuration, not from the card or the request. The exercise warns that A2A capabilities may be in preview.

How it fits

2.4.10 is the decision to compare: the exercise takes the tenant from a local setting and says production would derive it from verified identity claims. Ask which identity and connection 2.4.10 would require before this endpoint accepts delegated work.