This sample demonstrates a foundational ADK standalone agent that interacts with a user, manages session state via ToolContext, and uses multiple tools. Specifically, it features a hello_world_agent that can roll an N-sided die (storing roll history in the session state) and check whether numbers in a list are prime.

• Hi

  General greeting that does not trigger tool calls.

• Roll a dice with 100 sides

  The agent invokes the roll_die tool with sides=100. The rolled result is appended to the session's ToolContext state under the 'rolls' key.

• Roll a dice again with 100 sides.

  The agent invokes roll_die again, appending a second roll to the session state.

• What numbers did I got?

  The agent references the conversation history and previous tool outcomes to summarize the rolled numbers.

• Roll a die with 8 sides and check if the result is prime.

  Demonstrates multi-step tool orchestration. The agent first calls roll_die(sides=8), waits for the response, and then calls check_prime(nums=[...]) with the rolled result before formulating its final response.

graph TD
    Agent[Agent: hello_world_agent] --> Tool1[Tool: roll_die]
    Agent --> Tool2[Tool: check_prime]

Demonstrates how tools can access and modify persistent session state by including tool_context: ToolContext as a parameter.

def roll_die(sides: int, tool_context: ToolContext) -> int:
  result = random.randint(1, sides)
  if not 'rolls' in tool_context.state:
    tool_context.state['rolls'] = []
  tool_context.state['rolls'] = tool_context.state['rolls'] + [result]
  return result

Demonstrates adjusting GenerateContentConfig safety settings to prevent false alarms (e.g., avoiding harm category triggers when discussing rolling dice).

root_agent = Agent(
    model='gemini-3-flash-preview',
    name='hello_world_agent',
    ...
    generate_content_config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category=types.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
                threshold=types.HarmBlockThreshold.OFF,
            ),
        ]
    ),
)

You can execute the agent and inspect its session state programmatically by initializing an InMemoryRunner, creating a session, and executing prompts asynchronously:

runner = InMemoryRunner(agent=agent.root_agent, app_name='my_app')
session = await runner.session_service.create_session('my_app', 'user1')

async for event in runner.run_async(
    user_id='user1',
    session_id=session.id,
    new_message=types.Content(...),
):
  # Process execution events
  pass

# Inspect modified session state
session = await runner.session_service.get_session('my_app', 'user1', session.id)
print(session.state['rolls'])