A2A: Agent-to-Agent Protocol: interview questions

Production questions of the kind asked in applied AI engineering interviews.

1. Larkspur's support desk agent calls the handbook agent over A2A, then passes the reply to a writer agent. Customers start getting answers that quote context_id= and TaskStatus(. Nothing errored. Where do you look?

What a strong answer covers:

• Logs the raw JSON-RPC response from the handbook agent and checks result.kind: the agent now returns a Task, where it used to return a Message
• Finds the connector that reads result.parts and falls back to str(result) for a Task; the answer is in status.message (or in artifacts, if the agent publishes them)
• Fixes the reader to branch on kind, and fails loudly on a kind it does not know rather than stringifying it
• Adds a contract test that runs the client against both reply shapes, so a change on the server side is caught before customers see it

2. The platform team wants every Larkspur team to reach the handbook RAG system. One proposal is an MCP tool handbook_search; the other is an A2A agent larkspur_handbook_agent. Which do you ship?

What a strong answer covers:

• An MCP tool returns chunks and leaves reasoning to the caller's model; an A2A agent returns an answer from a system that owns retrieval, grading and rewriting
• Callers that need citations, their own prompt or their own reranker are better served by the tool; callers that want an answer and a task they can poll are better served by the agent
• A2A adds a task lifecycle, streaming and cancellation, and costs a second model loop per question
• The two can coexist: the agent behind A2A, with its retriever exposed as an MCP tool for callers that want raw chunks

3. The handbook agent runs behind A2A in production. What do you measure, and what alerts you that it is degrading?

What a strong answer covers:

• Counts tasks by final state (completed, failed, canceled, rejected) and JSON-RPC error codes per caller and per skill
• Measures time to the first working update and time to completed, separately from the graph's own latency
• Watches tasks stuck in working and tasks lost on restart, since an InMemoryTaskStore keeps nothing across a deploy
• Alerts when the Agent Card changes (version, skills, streaming) without a matching client release, and replays the golden questions through A2A after every deploy

4. Re-indexing the handbook takes 20 minutes, and the support desk triggers it through the handbook agent over message/send. Calls time out, and some re-index twice. How do you redesign the exchange?

What a strong answer covers:

• Has the client send with configuration.blocking set to false, which returns the Task at once in submitted, then poll tasks/get or subscribe with message/stream, instead of holding one HTTP request open for 20 minutes
• Uses push notifications (capabilities.push_notifications, tasks/pushNotificationConfig/set) when the client cannot hold a stream
• Makes the trigger idempotent per handbook version, so a retried send does not start a second re-index
• Persists tasks in a durable store and implements cancel(), since a cancel on a finished task answers -32002 and a lost task answers -32001

5. A partner company wants its booking agent to call Larkspur's handbook agent. The Agent Card is public. What do you check before you open it up?

What a strong answer covers:

• Declares authentication in the card (security_schemes, security) and keeps internal skills on the authenticated extended card (supports_authenticated_extended_card)
• Treats every incoming message as untrusted input: the partner's agent can carry prompt injection into the handbook agent's model
• Treats every reply from a remote agent as untrusted too, before it reaches a tool call or a customer
• Rate-limits and logs per caller, and scopes the skills a partner sees to what its contract covers