Session/State usage confusion

[ManasaMedepalli]

Hi Guys,
My agent treats each user message as an independent query instead of recognizing that a Yes/No response is feedback on the previous agent response—how do I implement state management to link multi-turn interactions?
For now, it is only two turn, FLOW: user -> root agent -> my Q agent -> answer back to root agent -> user feedback (Yes/No) and then I have to store this entire (each) iteration in db (this is also done in the Q agent). So basically want my root agent to just direct user questions to the agent, without any implementation in itself. HELP!
I'm a little confused around this concept and I posting a question here as I simultaneously also research it.
TIA :-)

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: You can manage the conversation state using session.state, a dictionary-like object that persists across multiple turns in the same conversation. To make your agent remember its last response, configure it with the output_key parameter.

---

Hi Manasa!

It sounds like you're running into a very common and important challenge in agent development: maintaining conversational context. ADK is designed to handle this through its Session and State management features.

Understanding Sessions and State

• Session: A Session in ADK represents a single, continuous conversation with a user. It tracks the entire history of interactions (user messages, agent responses, tool calls, etc.). [1]
• State (session.state): This is the key to solving your problem. session.state is a dictionary that acts as a temporary "scratchpad" or short-term memory for the current session. Any data you save to session.state in one turn can be read in subsequent turns, allowing your agent to remember what was said before. [2]

How to Solve Your "Yes/No" Problem

Here is a step-by-step approach to link your user's "Yes/No" feedback to the agent's previous response:

1. Save the Agent's Response to the State

You can configure your "Q agent" to automatically save its response to the session state by using the output_key parameter. This is the easiest way to keep track of the last thing the agent said.

For example, when you define your agent:

from adk.agents import LlmAgent

# Your "Q Agent"
q_agent = LlmAgent(
    # ... other parameters
    output_key="last_agent_response"
)

Now, after this agent runs, its final response will be stored in session.state["last_agent_response"].

2. Access the Previous Response in a Tool

Next, when the user provides feedback ("Yes" or "No"), you need a tool that can process this feedback. This tool can access the stored response from the state to understand the context.

To access the state, your tool function should accept a ToolContext object. [3]

from adk.tools import tool
from adk.tools.tool_context import ToolContext

@tool
def process_user_feedback(feedback: str, tool_context: ToolContext) -> str:
    """Processes the user's feedback based on the last agent response."""

    # Retrieve the last response from the session state
    last_response = tool_context.state.get("last_agent_response")

    if last_response is None:
        return "I'm sorry, I don't remember what I said before."

    if feedback.lower() == "yes":
        # User confirmed the previous response. Proceed with next action.
        # You can now store this confirmed information in your DB.
        return f"Great! I have recorded that you confirmed: '{last_response}'"
    elif feedback.lower() == "no":
        # User denied the previous response. Handle accordingly.
        return "My apologies. What would you like me to do instead?"
    else:
        return "I didn't understand the feedback."

Your root agent would then be configured to use this process_user_feedback tool when it detects a feedback-style response from the user.

State in a Multi-Agent Setup

Since you have a root agent and a Q agent, it's important to know that the session.state is shared across all agents operating within the same session. [4] When your root agent transfers control to the Q agent, the state persists, and when the conversation returns to the root agent, the state (including anything the Q agent added) is still there.

I hope this helps clear up your confusion! Let us know if you have more questions.

[1] https://google.github.io/adk-docs/get-started/about/
[2] https://google.github.io/adk-docs/sessions/
[3] https://google.github.io/adk-docs/tools-custom/
[4] https://google.github.io/adk-docs/tutorials/agent-team/

[ManasaMedepalli]

Hi Guys,
I am now trying out the FunctionTool (require_confirmation=true) feature to enable human in the loop.
it works fine on the root agent level, but if you have a subagent or an agent exposed as a tool and inside that, you have the function tool for any of its functions, the method doesn't work. the human feedback should go a level down from the root agent.
How can I achieve this? any input (also from humans , lol is appreciated )