How to implement human in the loop?

[qiankunli]

If the user's question is not clear enough, what kind of event will ADK throw? Where can the agent receive the user's confirmation to proceed with the execution?

[hangfei]

Hi @qiankunli , check out the doc here: https://google.github.io/adk-docs/tools/function-tools/#2-long-running-function-tool .

[kun321]

maybe this is you need https://google.github.io/adk-docs/agents/multi-agents/#human-in-the-loop-pattern

[pashna]

I'm curious about this as well. How would you implement the following scenario?

1. There's a "refund" tool that requires human approval before it can be executed.
2. To request approval, the agent initiates a call and then pauses (or "sleeps") until the approval response is received.
3. As approval might take a significant amount of time (e.g., days), the agent's state needs to be saved and cleared from memory during this waiting period,
4. Once the response is received, the agent is loaded again and continues the tool execution

Thanks!

[pashna]

Did anyone find the way?

[boyangsvl]

We are working on making the documents clearer. @polong-lin

This is what you do for human-in-the-loop:

1. You define a long running function and give it to the agent to kick off the human review. E.g. request_approval(...)
2. When needed, the agent will call that function.
3. In your function body, you write the code to kick off the work. E.g. send the approval request as a message to someone.
4. You can optionally return an intermediate result from your long running function, e.g. {'status': 'pending', 'eta': '2 days'}. The agent will use this info to formulate a response.
5. At this point the agent will finish executing because it knows that it needs to wait for the response of that long running function. You will need to record the function call id since you will need that to report the response.
6. minutes, hours, or days later, once the human approves, you call runner.run_async() again with the function id and function response. At this time the agent will continue to execute based on the function response.

[MahorShekhar]

@boyangsvl I also have the similar case where some function tool response will come later on. Right now I was handling with a kind of hack where I was registering event as author=system and role=model. I don't think this is the right way.

def handle_trigger(session: Session, trigger:str, form_data:dict = None) -> dict:

    state_delta = TRIGGER_MESSAGES[trigger]["state_delta"]

    system_event = Event(
        author="system",
        invocation_id=str(uuid.uuid4()),
        actions=EventActions(state_delta=state_delta),
        content=Content(role='model', parts=[Part(text=TRIGGER_MESSAGES[trigger]["event_content"])])
    )
    event = session_service.append_event(session, system_event)

    return {"event": event}

Now coming back to solution suggested by you, could you please help me in runner.run_async(), how can we provide function id and function response. Is there any example in cookbook ?