Running an agent

An agent is a description. A runner is the thing that drives it, and it needs one more object first: somewhere to keep the conversation.

Three names arrive together here and stay for the rest of the course. A session is one conversation. A session service stores sessions. A runner takes your agent, a session and a message, and drives the loop.

The agent, with the model from lesson 4

from pretend_adk import PretendModel, say, call
from google.adk.agents import LlmAgent

def lookup_order(order_id: str) -> str:
    """Look up the status of an order by its id."""
    return f"Order {order_id} shipped on 3 March."

The imports, and the tool from lesson 2 unchanged. Nothing here is new.

agent = LlmAgent(    name="support",
    model=PretendModel(replies=[call("lookup_order", order_id="A17"),
                                say("Your order shipped on 3 March.")]),
    instruction="Help customers with their orders.",
    tools=[lookup_order],
)

Two replies are scripted: first ask for the tool, then answer. That is one trip round the loop, written out in advance so the lesson can be about the runner rather than the model.

A runner and a session

from google.adk.runners import InMemoryRunner

runner = InMemoryRunner(agent=agent, app_name="demo")
session = await runner.session_service.create_session(app_name="demo", user_id="u1")

print("session id is a", type(session.id).__name__, "| events so far:", len(session.events))

InMemoryRunner is the runner plus a session service that keeps everything in memory. The session has an id, a history and a state, and the history is empty because nothing has happened yet.

Sending a message

A message is a content object with a role and parts, the same shape the model produces.

from google.genai import types

question = types.Content(role="user", parts=[types.Part(text="Where is order A17?")])

async for event in runner.run_async(user_id="u1", session_id=session.id, new_message=question):
    print(event.author, "->", "final" if event.is_final_response() else "step")

Three events for one question, all from the agent. The first two are the tool request and its result. The last is the answer, and is_final_response() is how you tell.

Getting just the answer

Most of the time that is all you want, so it is worth keeping these four lines.

session = await runner.session_service.create_session(app_name="demo", user_id="u1")

answer = ""
async for event in runner.run_async(user_id="u1", session_id=session.id, new_message=question):
    if event.is_final_response() and event.content:
        answer = "".join(part.text or "" for part in event.content.parts)

print(answer)

A fresh session, because the scripted model has used up its two replies. The loop keeps the text of whichever event says it is final, and that is your answer.