vertex-agent-engine-deployer

vertex-agent-engine-deployer

Deploy ADK 2.0 agents to Vertex AI Agent Engine — Google's managed runtime. No containers, no Kubernetes — Vertex hosts and scales it.

When to use

• You don't want to operate Cloud Run / GKE
• Want native Vertex session storage and observability
• Need built-in tracing, logging, scaling
• Multi-region / multi-tenant managed by Google

Deploy script

# deploy.py
import vertexai
from vertexai.preview import agent_engines
from agent import root_agent

PROJECT = "my-project"
LOCATION = "us-central1"
STAGING_BUCKET = "gs://my-adk-staging"

vertexai.init(project=PROJECT, location=LOCATION, staging_bucket=STAGING_BUCKET)

remote_agent = agent_engines.create(
    root_agent,
    requirements=[
        "google-adk[extensions] >= 2.0.0",
    ],
    display_name="my_adk_agent",
    description="ADK 2.0 agent deployed via Agent Engine.",
    extra_packages=["./agent.py"],
)

print(f"Resource name: {remote_agent.resource_name}")
# e.g. projects/.../locations/us-central1/reasoningEngines/1234567890

Run:

pip install google-cloud-aiplatform[agent_engines] vertexai
python deploy.py

Call the deployed agent

from vertexai.preview import agent_engines

remote_agent = agent_engines.get(
    "projects/PROJECT/locations/us-central1/reasoningEngines/1234567890"
)

session = remote_agent.create_session(user_id="u1")
response = remote_agent.run(session=session, input="Hello")
print(response)

Update an existing deployment

remote_agent = agent_engines.get(RESOURCE_NAME)
remote_agent.update(
    agent=updated_root_agent,
    requirements=[...],
)

IAM

The Agent Engine runtime needs:

• roles/aiplatform.user for itself (auto-assigned)
• Permissions for any backing services your tools call (DBs, APIs)

For caller authentication:

remote_agent.run(
    session=session,
    input="...",
    auth_token=user_jwt,  # propagated to tools
)

Pricing notes

Agent Engine bills for:

• Compute time per request
• Model tokens (Vertex AI standard rates)
• Session storage

Cheaper than always-on GKE for low-volume; comparable to Cloud Run.

Listing & cleanup

for a in agent_engines.list():
    print(a.resource_name, a.display_name)

agent_engines.delete(RESOURCE_NAME)

Validation

• agent_engines.list() shows the new deployment
• A run() call returns a non-empty response
• Cloud Trace shows agent spans
• Check Cloud Logging for any tool-call errors
• Cost dashboard reflects usage

See also

• cloud-run-deployer for full container control
• gke-deployer for max control / specialized hardware