Pass tool use from sub-agents in Agents-As-Tools pattern

[askurat]

I am looking to pass tool invocations from sub-agents. I want to pass the tool name and output that was used in the sub-agent down to my chatbot.

Currently, I only get the tool call which is just my sub-agent.

Any way to do this would be greatly appreciated.

[zastrowm]

There's nothing built-in to automatically do this for you, but in theory you could stream back all sub-events from the sub-agent back to the caller through tool-streaming: https://strandsagents.com/latest/documentation/docs/user-guide/concepts/tools/python-tools/#tool-streaming

[askurat]

Hi @zastrowm - thanks for the reply. I have played with this but this doesn't seem to accomplish what I am exactly trying to do. Or maybe I am doing it wrong.

Will it stream the whole response back up through my orchestrator? I want to be able to know what tool call was made within the sub-agent and it's output. On the UI side, I want have a custom component based on the tool used and it's output.

[zastrowm]

Will it stream the whole response back up through my orchestrator?

It should, yes, but it will be nested inside the tool_stream_event event. Do you have a code sample?

[askurat]

Here is some sample code. I tried to simplify it as much as possible.
Basically, the unified_orchestrator is using the task_create_agent as a tool. Within the task_create_agent, there is a create_task tool. This is streaming events.

I am trying to capture and send these events in the chat_router. The issue is the only tool call that shows is the task_create_agent not the create_task within the task_create_agent.

chat_router_test.py
main_test.py
task_create_agent.py
unified_orchestrator_test.py

[zastrowm]

In handle_task_creation, you have:

response = task_create_agent(context)

which is getting the agent's final response as a string; I think with what you're trying to accomplish, you'll need to yield the individual event from the task_create_agent, which looks more like this:

events = task_create_agent.invoke_async(context)
async for event in events:
     yield event
...

yield "Done" # this will become the tool's final result as seen by the LLM - all other events are not presented to the LLM

Then your orchestrator will start getting the events from the task_create_agent