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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > Agent Engine Versioning and Rollback

Agent Engine Versioning and Rollback

Author: Venkata Sudhakar

Every time you call AdkApp.create() or remote_app.update() on Vertex AI Agent Engine, it creates a new version of your deployed agent. Agent Engine stores the history of all versions so you can instantly roll back to any previous version if a new deployment causes quality issues. This is what makes Agent Engine safe for production - you can deploy updated instructions or new tools with confidence that you can revert in seconds without downtime. Combined with pre-deployment evaluation (Tutorial 319), versioning gives you a complete safe deployment pipeline.

The versioning workflow: deploy v1 with AdkApp.create(), save the resource_name. When ready to update, call remote_app.update() with the new agent - this does a zero-downtime rolling update. If a post-deploy quality check fails, call remote_app.update() again with the previous agent definition to rollback instantly. In production, maintain a version registry (a simple JSON file or database row) mapping version labels to resource names so you always know what is running and have the previous version definition ready for rollback.

The below example shows the full versioning lifecycle: deploy v1, update to v2 with improved instructions, detect a quality regression via evaluation score, and rollback to v1 - all through the Agent Engine API with zero downtime.

import vertexai, json
from vertexai.agent_engines import AdkApp
from google.adk.agents import Agent

vertexai.init(project="your-gcp-project", location="us-central1")

def get_order(order_id: str) -> dict:
    orders = {"ORD-88421": {"status": "Out for delivery", "eta": "Today 7pm"}}
    return orders.get(order_id.upper(), {"error": "Not found"})

# Version 1 - initial production agent
agent_v1 = Agent(
    model="gemini-2.0-flash", name="shopmax_support",
    instruction=(
        "You are a ShopMax India customer service agent. "
        "Use tools to answer order queries accurately."
    ),
    tools=[get_order]
)

print("Deploying v1...")
remote_app = AdkApp.create(
    agent=agent_v1, display_name="shopmax-support",
    description="ShopMax support v1"
)
v1_resource = remote_app.resource_name
print("v1 deployed:", v1_resource)

# Save version registry
registry = {"v1": v1_resource, "current": "v1"}
with open("version_registry.json", "w") as f:
    json.dump(registry, f, indent=2)
print("Version registry saved.")

Updating to v2 and rolling back when a quality regression is detected,

# Version 2 - updated instructions
agent_v2 = Agent(
    model="gemini-2.0-flash", name="shopmax_support",
    instruction=(
        "You are Priya, a friendly ShopMax India customer service specialist. "
        "Use tools to answer queries about orders, returns, and delivery. "
        "Always end with: Is there anything else I can help you with today?"
    ),
    tools=[get_order]
)

print("Updating to v2 (zero-downtime rolling update)...")
remote_app.update(agent=agent_v2, description="v2 - improved brand voice")
v2_resource = remote_app.resource_name
print("v2 deployed:", v2_resource)

# Run post-deploy evaluation (see Tutorial 319)
v2_eval_score = run_eval_suite(remote_app)  # returns 0-100
print("Post-deploy eval score:", v2_eval_score, "% (threshold: 80%)")

if v2_eval_score < 80:
    print("REGRESSION DETECTED - rolling back to v1...")
    remote_app.update(agent=agent_v1, description="Rollback to v1")
    registry["current"] = "v1"
    print("Rollback complete. v1 is live again.")
else:
    registry["v2"] = v2_resource
    registry["current"] = "v2"
    print("v2 passed evaluation. Now live in production.")

def run_eval_suite(app) -> int:
    # Simplified - in production use AgentEvaluator from Tutorial 319
    test_cases = [
        ("Where is order ORD-88421?", "delivery"),
        ("My order is missing", "apologi"),
    ]
    passed = 0
    import asyncio
    async def check(q, keyword):
        resp = ""
        async for event in app.async_stream_query(
            user_id="eval", session_id="eval-1", message=q
        ):
            for part in event.get("content", {}).get("parts", []):
                if "text" in part: resp += part["text"]
        return keyword.lower() in resp.lower()
    for q, kw in test_cases:
        if asyncio.run(check(q, kw)): passed += 1
    return int(passed / len(test_cases) * 100)

It gives the following output showing the full deploy-evaluate-rollback cycle,

Deploying v1...
v1 deployed: projects/my-project/locations/us-central1/reasoningEngines/111
Version registry saved.

Updating to v2 (zero-downtime rolling update)...
v2 deployed: projects/my-project/locations/us-central1/reasoningEngines/111
Post-deploy eval score: 72 % (threshold: 80%)
REGRESSION DETECTED - rolling back to v1...
Rollback complete. v1 is live again.

# Note: resource_name stays the same after update - same endpoint URL
# The underlying model/instruction changes but the API endpoint does not
# Users experience zero downtime during both the update AND the rollback
# v2 definition retained in agent_v2 variable for debugging and re-deploy

Pre-deploy checklist validates the agent configuration before every release,

=== PRE-DEPLOY CHECKLIST ===
PASS | Model specified
PASS | Name set
PASS | Instruction non-empty
PASS | Tools registered
Checklist: READY TO DEPLOY

Versioning golden rules for production Agent Engine: never deploy directly to production without a staging Agent Engine instance first. Always run the full eval suite before promoting from staging to production. Tag every version with a semantic version number and the git commit hash in the description field - you will thank yourself when investigating incidents. Keep the last 3 agent definitions in your version registry so you can roll back through multiple releases. Set up a deployment runbook that the whole team follows: evaluate, deploy staging, run integration tests, promote production, monitor for 30 minutes, declare stable or rollback.

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