Design suggestion for multi-agent architecture with streaming support

[stiven-hidri]

Hi! First of all, thanks for the great work on strands-agents.

I've recently implemented a single agent using the SDK, with access to a set of tools and support for streaming communication. As the complexity of my use case grows, I'm considering moving towards a multi-agent setup where agents are separated by domain.

The "agent-as-tool" pattern seemed promising, but during testing I noticed that when one agent calls another as a tool, streaming is no longer supported—the interaction becomes batch-like. This breaks a key requirement for my application, which relies heavily on real-time streaming responses.

Given this, I'd like to ask:
What is the recommended approach in the SDK for building a multi-agent system while preserving streaming communication?
Would a custom solution (e.g., a router agent that dynamically delegates tasks to the most appropriate agent) be more appropriate, or is there a pattern or capability in strands-agents that I might be overlooking? (e.g. streaming the agent-as-tool's response back to the orchestrator agent instead of returning it batch-like)

@tool
def research_assistant(query: str) -> str:
    """
    Process and respond to research-related queries.
 
    Args:
        query: A research question requiring factual information
 
    Returns:
        A detailed research answer with citations
    """
    try:
        research_agent = Agent(
            system_prompt=RESEARCH_ASSISTANT_PROMPT,
            tools=[retrieve, http_request]  # Research-specific tools
        )

        response = research_agent(query) # ***ENABLE STREMING RESPONSE***
        return str(response)
    except Exception as e:
        return f"Error in research assistant: {str(e)}"

Thanks in advance for any insights or recommendations!

[Alessandrus00]

Great point. It would be very useful if the output from the agent-as-tool could be streamed to the orchestrator in real time, rather than waiting for the entire response to be generated before the orchestrator can begin streaming it.