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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK with Cloud Tasks for Deferred Execution

ADK with Cloud Tasks for Deferred Execution

Author: Venkata Sudhakar

Some agent tasks are too slow to run inline with a user request - generating a 50-page report, processing a large data export, or running a multi-step pipeline. Cloud Tasks lets you enqueue these jobs and execute them asynchronously on a Cloud Run worker. ShopMax India uses Cloud Tasks to generate monthly sales analysis reports without blocking the dashboard UI.

The pattern: the main application enqueues a task with the job parameters, Cloud Tasks delivers it to a Cloud Run handler after an optional delay, and the worker runs the ADK agent and stores the result. The client polls or receives a notification when the job completes.

The below example shows how ShopMax India enqueues a report generation task and processes it asynchronously with an ADK agent.

from google.cloud import tasks_v2
import json
import base64

client = tasks_v2.CloudTasksClient()
QUEUE_PATH = client.queue_path("shopmax-india", "asia-south1", "report-jobs")

def enqueue_report_job(region: str, month: str, requester_email: str) -> str:
    payload = json.dumps({
        "region": region,
        "month": month,
        "requester": requester_email
    }).encode("utf-8")

task = {
        "http_request": {
            "http_method": tasks_v2.HttpMethod.POST,
            "url": "https://shopmax-agent-worker-xxxx.run.app/generate-report",
            "headers": {"Content-Type": "application/json"},
            "body": base64.b64encode(payload).decode("utf-8"),
            "oidc_token": {
                "service_account_email": "shopmax-agent-sa@shopmax-india.iam.gserviceaccount.com"
            }
        }
    }

response = client.create_task(parent=QUEUE_PATH, task=task)
    print(f"Task enqueued: {response.name}")
    return response.name

# Enqueue a report for Mumbai, October
task_name = enqueue_report_job("Mumbai", "2024-10", "manager@shopmax.in")

The Cloud Run worker that processes the task with an ADK agent,

from flask import Flask, request
from google.adk.agents import Agent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.adk.types import Content, Part
from google.cloud import storage
import json

app = Flask(__name__)

agent = Agent(
    name="report_agent",
    model="gemini-2.0-flash",
    description="Generates regional sales reports for ShopMax India"
)

@app.route("/generate-report", methods=["POST"])
def generate_report():
    job = request.get_json()
    region = job["region"]
    month = job["month"]

session_svc = InMemorySessionService()
    runner = Runner(agent=agent, session_service=session_svc, app_name="shopmax")
    sess = session_svc.create_session(app_name="shopmax", user_id="system")

prompt = f"Generate a sales performance report for ShopMax India {region} region for {month}."
    result = runner.run(session_id=sess.id, user_id="system",
                        new_message=Content(parts=[Part(text=prompt)]))
    report_text = ""
    for event in result:
        if event.is_final_response():
            report_text = event.content.parts[0].text

# Store result in Cloud Storage
    gcs = storage.Client()
    bucket = gcs.bucket("shopmax-reports")
    blob = bucket.blob(f"reports/{region}/{month}.txt")
    blob.upload_from_string(report_text)
    print(f"Report saved to GCS: reports/{region}/{month}.txt")
    return ("OK", 200)

It gives the following output,

Task enqueued: projects/shopmax-india/locations/asia-south1/queues/report-jobs/tasks/12345

# Worker logs when task executes:
Processing report job: Mumbai / 2024-10
Agent generated report: 847 words
Report saved to GCS: reports/Mumbai/2024-10.txt
Notification sent to: [email protected]

Cloud Tasks automatically retries failed tasks with exponential backoff. Set max_attempts to 3 and max_retry_duration to 24 hours for report jobs. Use the task name as a deduplication key to prevent duplicate reports if the same job is accidentally enqueued twice.

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