Is it possible to implement human in the loops in adk web?

[NBestsuNgan]

I know that there are example about this but it need to do with python app but i do want to know is it possible to implement it in adk web too?

[hangfei]

Here are related docs:

https://google.github.io/adk-docs/agents/multi-agents/#human-in-the-loop-pattern

https://google.github.io/adk-docs/tools/function-tools/#2-long-running-function-tool

@wyf7107 can share how it works on UI.

[NBestsuNgan]

@wyf7107 can you share some code to me please?

[wyf7107]

Are you trying to see how to approve things in adk web? We support long running functions like approvals in web. Or are you looking for something else?

[NBestsuNgan]

yes i want to see approve thing in ask web and i also try your example in the doc already but i still cannot imagine how to make an agent wait for my response, In case of sequential agent, the second agent will act immediately and i don't want that, I also try message = input("Approval or not: ") too but it ask for my approve in terminal instead:) but i do really want the agent to wait and ask me in web ui.

it fine if now we can't do that.

note that : nda_validate_bl_agent have no prompt

root_agent = SequentialAgent(
    name="adk_nda",
    sub_agents=[nda_main_agent, nda_validate_bl_agent],
    description="A pipeline that perform nda_main_agent, nda_validate_bl_agent, nda_type_agent, nda_filling_info_agent, nda_creation_agent respectively",
)

from .prompt import AGENT_INSTRUCTION
from google.adk.tools import LongRunningFunctionTool, FunctionTool
from google.adk.agents import Agent


MODEL = 'gemini-2.0-flash'
 
def ask_for_uderstand() -> str:
    """Ask for understanding of agent scope"""
    # create a ticket for the approval
    # Send a notification to the approver with the link of the ticket
    return {'status': 'pending', 'approver': 'User'}

    # message = input("Approval or not: ")
    # return {'status': message}

ask_for_approve_tool = LongRunningFunctionTool(func=ask_for_uderstand)

nda_main_agent = Agent(
    name="nda_main_agent",
    model=MODEL,
    instruction=AGENT_INSTRUCTION,
    tools=[ask_for_approve_tool],
    output_key="human_decision"

)

prompt

AGENT_INSTRUCTION="""You are a Main NDA task front-end user reception assistant.
** Instruction **
1. Greet the user with this message: 'Hi, I'm NDA agent. I'm restricted to creating NDA documents only. I will now proceed to start the process.'
2. Immediately call `ask_for_approve_tool` to get user approval by waiting for user input.
"""