How to Pass State in async_stream_query

[RafayJ-XPO]

I have created a simple agent and then deployed it to agent engine, however I am unable to pass the state as a param to input in async_stream_query class method. In my local development using adk web I can simply pass the properties and state is updated.

How can this be done when the agent in deployed to agent engine

[adk-bot]

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

TLDR: When an agent is deployed to Agent Engine, you do not pass the state directly into the async_stream_query method. Instead, state is automatically managed on the backend for each conversation (session). You access and modify the state from within your agent's tools and logic.

---

Hello! That's a great question. The way state is handled shifts when you move from local development to a deployed Agent Engine environment, which can be a bit confusing.

How State Works in Agent Engine

When you deploy your ADK application to Agent Engine, it automatically uses a persistent SessionService to manage conversational state. [1] This means for each unique conversation or session, Agent Engine maintains a state object on the backend. This state persists across multiple turns of the same conversation, giving your agent a "memory" for that specific session. [2]

Because the state is managed server-side and tied to a session_id, you don't (and can't) pass it as a parameter in your client-side call to async_stream_query.

How to Access and Modify State

The correct way to interact with the session state is from within your agent's internal logic, primarily through its tools.

1. Using Tools: You can design your tools to automatically receive a ToolContext object. This object contains a state attribute, which is a dictionary-like object representing the current session's state. Your tool can then read from or write to this dictionary to manage the conversational state. [3, 4]

   Here is a simple example of a tool that accesses state:

   from adk.tools import tool, ToolContext
   
   @tool
   def remember_name(name: str, context: ToolContext) -> str:
       """Remembers the user's name for the rest of the conversation."""
       context.state["user_name"] = name
       return f"Got it! I'll remember your name is {name}."
   
   @tool
   def recall_name(context: ToolContext) -> str:
       """Recalls the user's name that was previously remembered."""
       name = context.state.get("user_name", "friend")
       return f"Hello, {name}!"

2. Using Agent Output: An agent can also be configured to automatically save its final output to the state by setting the output_key property. This is useful for chaining agents together in a workflow, where one agent's output becomes the input for another via the shared state. [5]

In summary, instead of trying to push state to the agent with your query, you should be passing a unique session_id for each conversation. Agent Engine will handle loading the correct state for that session, and your agent's internal tools and logic can then access and modify it as needed. [2]

[1] https://google.github.io/adk-docs/deploy/agent-engine/
[2] https://google.github.io/adk-docs/sessions/
[3] https://google.github.io/adk-docs/tools-custom/
[4] https://google.github.io/adk-docs/sessions/state/
[5] https://google.github.io/adk-docs/tutorials/agent-team/

[RafayJ-XPO]

How to pass extra parameter or data to any session