How to maintain session context and state persistence when using sub-agents via AgentTool?

[Campeoni]

Hi everyone,

I’m working with the ADK TypeScript SDK and I’ve run into a behavior regarding state persistence that I’m trying to better understand.

The Scenario:
I have a parent agent (Agent A) that uses a sub-agent (Agent B) as a tool via AgentTool. Agent B is configured with an outputKey to persist its results into the session state.

What I'm observing:

When I run Agent B standalone, it works perfectly: the result is saved to the outputKey in the session state.

When Agent B is called as a tool by Agent A, it executes and returns the correct response to the parent, but the outputKey in the state is not updated.

It seems like the sub-agent might be running in an isolated or temporary session context (I've noticed some references to temp_user in logs), which prevents it from writing to the "real" parent session state.

My Questions:

Is it the expected behavior that sub-agents in AgentTool have an isolated state?

Is there a specific configuration required to "pass through" or inherit the sessionId and userId from the parent agent to the sub-agent so they share the same state?

Am I perhaps missing a step in how the Runner handles state updates from nested agents?

I’ve tried using afterModel callbacks to manually set the state, but while the state object changes during execution, those changes don't seem to persist back to the main session once the tool finishes.

Any guidance on the "canonical" way to handle shared state in multi-agent architectures would be greatly appreciated!