How to handle long running agent with A2A?

[nithin-nk]

We have LangGraph agents, which will take 10 to 15 minutes to run. How to handle long-running agents with A2A. Is it possible to maintain a live connection for 15 minutes? Is there any background task execution? Is there a better way to do it?

[RubberGoose246]

I think the key concept of A2A is task-based communication between remote agents, and it’s not meant to be limited to chat-style interactions where no real task management is needed and everything finishes quickly. However, most of the current sample implementations focus heavily on LLM-based chat agent scenarios, and the documentation and examples are lacking when it comes to long-running tasks and broader task lifecycle management.

[lukehinds]

You would use Push notifications.

Tasks in A2A evolve through a set of statuses:

a2a-python/src/a2a/types.py

Lines 908 to 916 in d9e463c

	submitted = 'submitted'
	working = 'working'
	input_required = 'input-required'
	completed = 'completed'
	canceled = 'canceled'
	failed = 'failed'
	rejected = 'rejected'
	auth_required = 'auth-required'
	unknown = 'unknown'

You can poll these for completed, or use a push notification (webhook) for the Agent to callback when it completes its task.

[ahoblitz]

https://a2a-protocol.org/latest/topics/life-of-a-task/

[pstephengoogle]

Yes, A2A supports polling, pushes or streams.

For a true long running job where keeping the network connection open is not feasible, the best approach is to send the request with blocking=False. This will return the task information immediately and can be treated as a long running operation from there. You can provide a push notification for the agent to send updates to, you can create your own poller and call tasks/get on whatever cadence you feel is appropriate, or you can call tasks/subscribe to reestablish a new streaming connection for new events.

Will consider examples illustrating these patterns