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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK with Cloud Workflows for Orchestration

ADK with Cloud Workflows for Orchestration

Author: Venkata Sudhakar

Google Cloud Workflows provides a managed orchestration service that complements ADK agents by handling long-running processes, retry logic, and multi-service coordination. When an ADK agent needs to coordinate across multiple systems or run tasks that span hours, Cloud Workflows acts as the durable execution layer while ADK provides the intelligence.

ShopMax India uses this combination to automate supplier onboarding - a process that involves document verification, credit checks, catalogue ingestion, and system provisioning across five different services. The ADK agent handles all the intelligent decisions while Cloud Workflows manages the overall process durability.

The below example shows how to define a Cloud Workflows YAML definition that calls an ADK agent endpoint at each decision step.

# cloud_workflow.yaml - ShopMax Supplier Onboarding Workflow
main:
  params: [args]
  steps:
    - validate_documents:
        call: http.post
        args:
          url: ${sys.get_env("AGENT_ENDPOINT") + "/validate"}
          body:
            supplier_id: ${args.supplier_id}
            documents: ${args.documents}
        result: validation_result
    - check_validation:
        switch:
          - condition: ${validation_result.body.status == "approved"}
            next: credit_check
          - condition: ${validation_result.body.status == "rejected"}
            next: notify_rejection
    - credit_check:
        call: http.post
        args:
          url: ${sys.get_env("AGENT_ENDPOINT") + "/credit"}
          body:
            supplier_id: ${args.supplier_id}
        result: credit_result
        retry:
          predicate: ${default_retry_predicate}
          max_retries: 3
          backoff:
            initial_delay: 2
            max_delay: 60
            multiplier: 2
    - notify_rejection:
        call: http.post
        args:
          url: ${sys.get_env("NOTIFY_ENDPOINT")}
          body:
            message: "Supplier application rejected"
            supplier_id: ${args.supplier_id}
        next: end
    - end:
        return: ${credit_result.body}

It gives the following output,

Workflow execution started: exec-abc123
Step validate_documents: POST /validate -> status=approved
Step credit_check: POST /credit -> credit_score=720, limit=Rs 500000
Workflow completed successfully in 4.2s

The below example shows the ADK agent that handles the intelligent decisions at each workflow step.

import functions_framework
from google.adk.agents import LlmAgent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
import vertexai
import json

vertexai.init(project="shopmax-india", location="asia-south1")

def create_validation_agent():
    return LlmAgent(
        model="gemini-2.0-flash",
        name="supplier_validator",
        instruction=(
            "You are a supplier document validator for ShopMax India. "
            "Analyse submitted documents and return approved or rejected "
            "with specific reasons. Check for: GST registration, PAN card, "
            "bank details, and address proof."
        )
    )

@functions_framework.http
def validate(request):
    data = request.get_json()
    supplier_id = data["supplier_id"]
    documents = data["documents"]

agent = create_validation_agent()
    session_service = InMemorySessionService()
    runner = Runner(
        agent=agent,
        app_name="supplier_onboarding",
        session_service=session_service
    )

session = session_service.create_session(
        app_name="supplier_onboarding",
        user_id=supplier_id
    )

prompt = f"Validate documents for supplier {supplier_id}: {json.dumps(documents)}"
    result = None
    for event in runner.run(
        user_id=supplier_id,
        session_id=session.id,
        new_message={"role": "user", "parts": [{"text": prompt}]}
    ):
        if event.is_final_response():
            result = event.content.parts[0].text

status = "approved" if "approved" in result.lower() else "rejected"
    return {"status": status, "details": result, "supplier_id": supplier_id}

It gives the following output,

POST /validate
{"supplier_id": "SUP-2847", "documents": {"gst": "29AABCS1429B1ZB", "pan": "AABCS1429B"}}

Response:
{"status": "approved", "supplier_id": "SUP-2847",
 "details": "All documents verified. GST active, PAN matched, bank details confirmed."}

Cloud Workflows handles all retry logic, timeouts, and error propagation automatically, while the ADK agent focuses purely on the business decision. This separation means you can update the agent logic independently without changing the orchestration layer - making the system easier to maintain and test in a production environment like ShopMax India.

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