RU
Get in touch
Chapter 28

Choose Polling, Streaming, or Push Deliberately

A2A supports three mechanisms for delivering task updates, and the choice among them is an engineering decision, not a matter of taste. Polling is the simplest. Streaming is faster for interactive UX. Push fits very long server-to-server tasks. And whichever mechanism you choose must be declared in the Agent Card - the client has no right to assume what the agent didn't promise.

A comparison by cost and reliability helps you not to err.

Mechanism · When to take it · Cost · Reliability

  • Get Task polling - Rare updates, a strict firewall; Extra requests and latency; Recovers simply
  • Streaming/SSE - An interactive UI, frequent events; A long connection and reconnect logic; After a disconnect some messages may be lost
  • Push webhook - Very long server-to-server work; A public receiver, retries, verification; Needs idempotency and a delivery policy

A multi-turn scenario - when the agent asked for input and continued after the answer - looks like this at the wire level.

SQL
1. SendMessage → Task(input_required)
2. Client shows status.message
3. Client sends a new Message
   taskId = existing task
   contextId = server-issued context
4. Agent resumes → working
5. Artifact update → completed

And cancelling a task goes through an ownership check and doesn't undo already committed downstream effects.

client CancelTask(taskId)
server verifies task ownership
server rejects terminal/non-cancelable tasks
executor observes cancellation
server publishes canceled terminal state
already committed downstream effects require reconciliation

Two caveats here are especially costly. The first is about streaming.

Streaming messages are not a durable journal. The specification warns: after a disconnect the client may not receive all status messages. Keep a critical result in the Task status/history/artifacts and re-fetch it via Get Task, rather than relying on the stream.

The second is about push, which opens a new surface outward.

Push opens an SSRF surface. Check the callback URL, block private and link-local ranges, use HTTPS, sign callbacks, limit redirects, and re-resolve DNS by a safe policy.

The delivery is chosen. Now let's tie together what passes through all these mechanisms - security, tasks, and tenant ownership.

Links