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

	Facebook
	Twitter
	Linkedin
	E-Mail

Agentic AI > ADK Agent Testing > Testing ADK Agent Graceful Degradation When Tools Are Unavailable

Testing ADK Agent Graceful Degradation When Tools Are Unavailable

Author: Venkata Sudhakar

Graceful degradation is the ability of a system to continue operating in a reduced capacity when one or more of its components fail. For ShopMax India's ADK agents, this means the agent should still serve customers even when tools like inventory lookup, payment gateway, or shipping API become unavailable. Testing graceful degradation ensures the agent handles tool failures without crashing or returning unhelpful errors.

ADK agents implement graceful degradation through fallback strategies: returning cached data, using alternative tools, providing partial responses, or gracefully informing users of limited service. The key is that the agent should degrade predictably and transparently. Testing focuses on verifying that each degraded mode activates correctly and that users receive meaningful responses rather than raw exceptions.

The following example tests ShopMax India's product inquiry agent when the inventory tool and pricing tool are unavailable. It verifies the agent falls back to cached catalog data and informs users about limited availability:

import pytest
from unittest.mock import patch, MagicMock
from google.adk.agents import Agent
from google.adk.tools import FunctionTool
from google.adk.runners import InMemoryRunner

class ToolUnavailableError(Exception):
    pass

CACHED_CATALOG = {
    "SM-TV-55": {"name": "Samsung 55-inch 4K TV", "base_price": 45000},
    "LG-AC-15": {"name": "LG 1.5 Ton AC", "base_price": 38000},
    "PHI-WM-7": {"name": "Philips 7kg Washing Machine", "base_price": 22000},
}

def check_inventory(product_id: str) -> dict:
    raise ToolUnavailableError("Inventory service is down for maintenance")

def get_live_pricing(product_id: str) -> dict:
    raise ToolUnavailableError("Pricing service temporarily unavailable")

def get_catalog_fallback(product_id: str) -> dict:
    if product_id in CACHED_CATALOG:
        item = CACHED_CATALOG[product_id]
        return {
            "product_id": product_id,
            "name": item["name"],
            "base_price": item["base_price"],
            "note": "Live inventory unavailable. Showing catalog data. Call 1800-shopmax for stock."
        }
    return {"error": f"Product {product_id} not found in catalog"}

product_agent = Agent(
    name="product_inquiry_agent",
    model="gemini-2.0-flash",
    instruction=(
        "You are ShopMax India product assistant. When inventory or pricing tools fail, "
        "use catalog_fallback to provide basic product info and tell the customer to "
        "contact 1800-shopmax for live stock availability."
    ),
    tools=[
        FunctionTool(check_inventory),
        FunctionTool(get_live_pricing),
        FunctionTool(get_catalog_fallback),
    ]
)

runner = InMemoryRunner(agent=product_agent)
session = runner.session_service.create_session(
    app_name=runner.app_name, user_id="customer_mumbai"
)

def test_graceful_degradation_inventory_unavailable():
    """Agent uses catalog fallback when inventory tool raises ToolUnavailableError."""
    response = runner.run(
        user_id="customer_mumbai",
        session_id=session.id,
        message="Is Samsung 55-inch TV SM-TV-55 in stock? What is the price?"
    )
    reply = " ".join(e.content.parts[0].text for e in response
                     if e.content and e.content.role == "model")
    assert "catalog" in reply.lower() or "1800-shopmax" in reply.lower(), \
        "Agent should mention catalog data or helpline when tools are unavailable"
    assert "45000" in reply or "45,000" in reply, \
        "Catalog base price should appear in degraded response"
    assert "ToolUnavailableError" not in reply, \
        "Raw exception names must not leak into customer-facing response"

def test_graceful_degradation_all_tools_unavailable():
    """Agent provides meaningful response even when all tools fail."""
    with patch("__main__.get_catalog_fallback",
               side_effect=ToolUnavailableError("All systems down")):
        response = runner.run(
            user_id="customer_mumbai",
            session_id=session.id,
            message="Check stock for LG-AC-15 air conditioner"
        )
        reply = " ".join(e.content.parts[0].text for e in response
                         if e.content and e.content.role == "model")
        assert len(reply) > 0, "Agent must always return a response"
        assert any(word in reply.lower() for word in
                   ["unavailable", "sorry", "unable", "contact", "try again"]), \
            "Agent should express inability and suggest alternative action"

def test_degraded_response_contains_no_technical_details():
    """Degraded responses must be customer-friendly with no stack traces."""
    response = runner.run(
        user_id="customer_mumbai",
        session_id=session.id,
        message="Get price for Philips Washing Machine PHI-WM-7"
    )
    reply = " ".join(e.content.parts[0].text for e in response
                     if e.content and e.content.role == "model")
    forbidden_terms = ["traceback", "exception", "error:", "line ", "file ""]
    for term in forbidden_terms:
        assert term.lower() not in reply.lower(), \
            f"Technical term '{term}' must not appear in customer response"
    print(f"Degraded response sample: {reply[:200]}")

It gives the following output,

test_graceful_degradation_inventory_unavailable PASSED
test_graceful_degradation_all_tools_unavailable PASSED
test_degraded_response_contains_no_technical_details PASSED

3 passed in 6.41s

Degraded response sample: I was unable to retrieve live stock and pricing information for
the Philips 7kg Washing Machine (PHI-WM-7) as our systems are currently under maintenance.
Based on our catalog, the base price is Rs 22,000. For current stock availability,
please call 1800-shopmax or visit your nearest ShopMax India store in Mumbai or Bangalore.

Production guidance for ShopMax India: define a fallback priority chain - live inventory first, then cached catalog, then a generic apology message with helpline number. Monitor the degraded mode activation rate in Grafana; a spike means a downstream outage, not a test failure. Always test the fallback chain with real tool timeout values, not just immediate errors, since network timeouts are the most common real-world failure mode.

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