	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Agentic AI > ADK Agent Testing > State Transition Testing for Multi-Turn ADK Agents

State Transition Testing for Multi-Turn ADK Agents

Author: Venkata Sudhakar

State transition testing verifies that a multi-turn ADK agent moves through its conversation states correctly - accepting valid transitions and rejecting invalid ones. ShopMax India's order support agent has defined states (greeting, intent_identified, tool_called, response_given, session_closed) and must never skip a required state or re-enter a closed session, properties that are only caught by explicitly testing each transition path.

The test model is a state machine where each state has a set of valid next states. A StateTracker records the sequence of states visited during a conversation turn. Tests drive the agent through a valid path and assert the recorded sequence matches the expected state chain. Invalid transitions - like jumping from greeting directly to response_given - must raise a StateError or be blocked by the tracker.

The example below defines an OrderSupportStateMachine with five states, tests a valid end-to-end conversation path, and tests that an invalid transition from greeting to response_given is correctly rejected.

import pytest
from dataclasses import dataclass, field
from typing import List

VALID_TRANSITIONS = {
    "greeting":          {"intent_identified"},
    "intent_identified": {"tool_called"},
    "tool_called":       {"response_given"},
    "response_given":    {"greeting", "session_closed"},
    "session_closed":    set(),
}

class StateError(Exception):
    pass

@dataclass
class StateMachine:
    state: str = "greeting"
    history: List[str] = field(default_factory=lambda: ["greeting"])

def transition(self, next_state: str):
        allowed = VALID_TRANSITIONS.get(self.state, set())
        if next_state not in allowed:
            raise StateError(
                f"Invalid transition: {self.state} -> {next_state}. Allowed: {allowed}"
            )
        self.state = next_state
        self.history.append(next_state)

def test_valid_order_support_conversation():
    sm = StateMachine()
    sm.transition("intent_identified")
    sm.transition("tool_called")
    sm.transition("response_given")
    sm.transition("session_closed")
    expected = ["greeting", "intent_identified", "tool_called", "response_given", "session_closed"]
    assert sm.history == expected
    print(f"State path: {sm.history}")

def test_invalid_transition_rejected():
    sm = StateMachine()
    with pytest.raises(StateError, match="Invalid transition"):
        sm.transition("response_given")
    print(f"Invalid transition greeting->response_given correctly blocked")

def test_closed_session_blocks_reentry():
    sm = StateMachine()
    for state in ["intent_identified", "tool_called", "response_given", "session_closed"]:
        sm.transition(state)
    with pytest.raises(StateError):
        sm.transition("greeting")
    print("Closed session correctly blocked re-entry")

It gives the following output,

State path: ['greeting', 'intent_identified', 'tool_called', 'response_given', 'session_closed']
Invalid transition greeting->response_given correctly blocked
Closed session correctly blocked re-entry
3 passed in 0.05s

Draw the state machine diagram in the test file as a comment table so that new developers can visually verify test coverage against the spec. Test every valid path (not just the happy path) and every invalid transition from every state - this matrix approach ensures no edge in the state graph is untested. For agents with long conversation flows, use itertools.product to generate all transition combinations automatically rather than writing them by hand.

Send your comments, suggestions or queries regarding this site to [email protected].