Agent Development Kit : HIL and calling tools inside custom agents

[ashikaumanga]

I am building a Custom Agent using Google’s ADK as explained here : https://google.github.io/adk-docs/agents/custom-agents/

I've wrapped two agents as sub-agents within this custom agent.

I have a couple of questions:

1) User Input (HIL):
 Inside the custom agent, I want to get input from the user (Human-in-the-Loop). For now, I'm using input().
What is the recommended or best practice for handling user input in this context? Do I need to send some Events for this?

2) Calling a Tool (e.g., SQL execution):
I want the custom agent to call an external tool — for example, executing SQL on a database.
What is the best practice for integrating such tools inside a custom agent? Do I need to use Events for this ?

Code :

class IntentionFlowAgent(BaseAgent):
    
    model_config = {"arbitrary_types_allowed": True}
    sub_agent1: LlmAgent
    sub_agent2 : LlmAgent

    
    def agent_event(self, author: str, message:str, invocation_id) -> Event:
        
        return Event(author=author,
                     invocation_id=invocation_id,
                     content= { "parts": [ {"text": message }] } )

    def __init__(self,
                 name : str,
                 sub_agent1: LlmAgent,
                 sub_agent2: LlmAgent):
        super().__init__(name=name,
                         sub_agent1= sub_agent1,
                         sub_agent2= sub_agent2)

    async def _run_async_impl(self, ctx: InvocationContext) -> AsyncGenerator[Event, None]:

        while True:

            # call sub_agent1
            async for event in self.sub_agent1.run_async(ctx):
                yield event
 
            missing_domains = ctx.session.state.get("missing_domains")
            if missing_domains is not None:
                yield self.agent_event(self.name,"Missing domains found " +str(missing_domains) + " cant proceed further", ctx.invocation_id)
                return
            


            # call sub_agent2
            async for event in self.intention_refine_agent.run_async(ctx):
                yield event


            # check result generated by sub_agent2
            result : dict = json.loads(ctx.session.state["sub_agent2_result"])
            if  "intent" in result:
                ctx.session.state["final_user_intent"] = result['intent']
                break
            
            # Human In The Loop for clarifications
            ambiguity = result['ambiguity']
            user_feedback = input("-- Found ambiguity " + ambiguity + " : ")
            
            refinement_context = "\n\nQuestion : `" + ambiguity + "`\n\n Answer : `" + user_feedback + "`"
            
            # update the state
            tmp = ctx.session.state.get("prev_refinements" , list())
            tmp.append(refinement_context)
            ctx.session.state["prev_refinements"] = tmp

[drtn25]

yup i think so