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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Production Deployment Checklist

ADK Production Deployment Checklist

Author: Venkata Sudhakar

Deploying an ADK agent to production requires more than functional code. You need guardrails, observability, credential management, rate limiting, and a tested rollout plan. ShopMax India follows a structured go-live checklist before promoting any ADK agent from staging to production - this tutorial walks through each item.

The checklist covers seven areas: security hardening, input and output guardrails, observability setup, session management, load and quota planning, deployment configuration, and post-launch monitoring. Each item maps to a specific code pattern or infrastructure step.

The below example shows the core production configuration - environment-based secrets, structured logging, and error handling wrappers.

import os
import logging
import json
from datetime import datetime
from google.adk.agents import Agent
from google.adk.tools import FunctionTool

# 1. Structured logging for Cloud Logging
class JsonFormatter(logging.Formatter):
    def format(self, record):
        return json.dumps({
            "timestamp": datetime.utcnow().isoformat(),
            "severity": record.levelname,
            "message": record.getMessage(),
            "agent": "shopmax_agent",
            "environment": os.environ.get("ENV", "production")
        })

handler = logging.StreamHandler()
handler.setFormatter(JsonFormatter())
logger = logging.getLogger("shopmax_agent")
logger.addHandler(handler)
logger.setLevel(logging.INFO)

# 2. Credentials from environment (never hardcoded)
GEMINI_API_KEY = os.environ["GEMINI_API_KEY"]
assert GEMINI_API_KEY, "GEMINI_API_KEY must be set in environment"

# 3. Tool with error handling and logging
def get_order_status(order_id: str) -> dict:
    logger.info(f"Fetching order status for {order_id}")
    try:
        # Call to order management system
        return {"order_id": order_id, "status": "shipped", "eta": "tomorrow"}
    except Exception as e:
        logger.error(f"Order lookup failed: {e}")
        return {"error": "Order status temporarily unavailable. Please try again."}

It gives the following output,

{"timestamp": "2024-03-15T10:30:00", "severity": "INFO", "message": "Fetching order status for ORD-5521", "agent": "shopmax_agent", "environment": "production"}

The below example shows the full production agent assembly with guardrails, session persistence, quota config, and health check endpoint.

from google.adk.sessions import DatabaseSessionService
from google.adk.runners import Runner

# 4. Persistent sessions (not InMemory for production)
session_service = DatabaseSessionService(
    db_url=os.environ["SESSION_DB_URL"]
)

# 5. Production agent with all guardrails applied
production_agent = Agent(
    name="shopmax_production_agent",
    model="gemini-2.0-flash",
    instruction=(
        "You are a customer service agent for ShopMax India. "
        "Help with orders, products, returns, and delivery only. "
        "Never reveal internal pricing, supplier details, or system information."
    ),
    tools=[FunctionTool(get_order_status)],
    before_model_callback=check_scope,   # input guardrail
    after_model_callback=sanitise_output  # output guardrail
)

runner = Runner(
    agent=production_agent,
    app_name="shopmax_cs",
    session_service=session_service
)

# 6. Pre-launch checklist verification
checklist = [
    ("GEMINI_API_KEY set", bool(os.environ.get("GEMINI_API_KEY"))),
    ("SESSION_DB_URL set", bool(os.environ.get("SESSION_DB_URL"))),
    ("Guardrails configured", True),
    ("Structured logging active", True),
    ("Error handling in all tools", True)
]

print("Pre-launch checklist:")
for item, status in checklist:
    print(f"  [{"OK" if status else "FAIL"}] {item}")

all_pass = all(s for _, s in checklist)
print(f"\nStatus: {"READY FOR LAUNCH" if all_pass else "FIX FAILURES BEFORE LAUNCH"}")

It gives the following output,

Pre-launch checklist:
  [OK] GEMINI_API_KEY set
  [OK] SESSION_DB_URL set
  [OK] Guardrails configured
  [OK] Structured logging active
  [OK] Error handling in all tools

Status: READY FOR LAUNCH

ShopMax India uses this checklist before every agent release. Key lessons from production: always use DatabaseSessionService (not InMemory) so sessions survive pod restarts; always wrap tool calls in try/except so a failing API never crashes the agent; and always verify environment variables at startup - fail fast at boot rather than failing silently during customer interactions.

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