ADK fails to deploy with 1.18.0 and above

I can deploy agent to Vertex AI using adk in 1.17.0 with this command:

    --project=$PROJECT_ID \
    --region=$LOCATION \
    --staging_bucket=$STAGING_BUCKET \
    --display_name="$AGENT_NAME" \
    ArchitectureAgent

However if I update to 1.18.0 and above

pip install google-adk==1.18.0

deployment will fail with this message:

Deploy failed: Failed to create Agent Engine: {'code': 3, 'message': 'Reasoning Engine resource [projects/{redacted_project_ID}/locations/europe-west2/reasoningEngines/{redacted_agent_ID}] failed to start and cannot serve traffic. Please refer to our documentation (https://cloud.google.com/vertex-ai/generative-ai/docs/agent-engine/troubleshooting/deploy) for checking logs and other troubleshooting tips.'}

Still I have other issues with running agent even 1.17 that might or might not be related, but my guess is it shouldn't influence deployment.

We hit the same Agent Engine deployment failure (Reasoning Engine failed to start) and traced it to invalid import/module paths caused by the temp folder name and template import path. We’ve opened a PR with fixes for temp folder sanitization and proper template imports when the agent lives under app/: #4243. Hope this helps—happy to share more repro details if useful.

@Alameyo

Hello @Alameyo ,
We tried to reproduce this issue. Although no specific error message appears, the deployment process seems to be stuck for a long time.

Thank you, @smwitkowski for the PR. @ryanaiagent could you please provide guidance on the next steps required to merge this?

In the meantime, you can try wrapping your agent in agent_engines.AdkApp() and deploying it using agent_engines.create().

Below is a sample code snippet for your reference to help you deploy your agent:

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

# Initialize the agent
my_agent = Agent(
    model="gemini-2.5-flash",
    name="my_agent",
    description="Add two numbers and provide step-by-step reasoning for a child to understand.",
    instruction="You are a helpful assistant that can add two numbers and provide step-by-step reasoning for a child to understand it."
)

# Wrap the agent into AdkApp
app = agent_engines.AdkApp(
    agent=my_agent,
    enable_tracing=True,
)

# Deploy to Agent Engine
remote_app = agent_engines.create(
    agent_engine=app,
    requirements=[
        "google-cloud-aiplatform[adk,agent_engines]"   
    ]
)

print(f"Deployment finished!")
print(f"Resource Name: {remote_app.resource_name}")

Tried to deploy using adk deploy agent_engine, also this solution above using vertexai, and not worked (got the same 10 error code)

Python version: 3.12
Google-adk version 1.24.0

Leaved related issue on #4683

@francamacdowell

Hello @francamacdowell ,

We tested this on our end but everything seems to be working as expected. Could you please try updating to google-adk==1.26.0? While this should also work with version 1.24.0, updating might fix the problem. If it is still not working kindly share your code so we can take a closer look.

This issue has been automatically marked as stale because it has not had recent activity for 7 days after a maintainer requested clarification. It will be closed if no further activity occurs within 7 days.

On my side I just staid for my deployment with older version: 1.17 whch works for me, but I had separate issues running the agent on europe-west2. When I moved to us-central everything started working fine. I wonder if europe-west is missing something.

@Alameyo

Hello @Alameyo ,

Thank you for your inputs, we also tried this with us-central1 which worked fine.

Kindly refer to Vertex AI docs for detailed information.

My agent doesn't need to execute the code, so it should work fine in the europe-west2.

@yeesian

Hello @yeesian ,

Could you please share your input here? europe-west2 is not working but ideally it should work according to the documentation.

Adding another affected region: us-west1.

Same failure as reported — code: 3 / 400 "failed to start and cannot serve traffic". What may help narrow it down: both Agent Engine deployment modes fail in us-west1, while identical code deploys and serves fine in us-east1.

sourceCodeSpec (inlineSource + imageSpec — container built from a Dockerfile, via agents-cli deploy) — fails
packageSpec (pickleObjectGcsUri + requirementsGcsUri — plain agent_engines.create() with an AdkApp) — also fails

Since the second path never goes through cli_deploy.py, this looks like it sits below the deployment mode — i.e. regional runtime provisioning rather than the CLI. That would also mean PR #4243 (temp-folder sanitization / template import paths) fixes a different cause of the same symptom.

Diagnostic detail that may be the most useful part: in every us-west1 failure, Cloud Logging contains only aiplatform.googleapis.com/reasoning_engine_build entries — zero reasoning_engine_stderr and zero reasoning_engine_stdout. A successful us-east1 deploy of the same image emits stderr from the very first line (uv output at process start, since the container CMD is uv run uvicorn ...). So in us-west1 the container/runtime never executes at all; the failure precedes any user code.

Ruled out on our side:

Engine count / regional capacity — deleted an existing engine in the region, no change.
The Agent Engine runtime service account lacking logging/trace roles — us-east1 succeeds with the same SA and telemetry enabled.
OOM / resource limits — peak ~150 MiB against the 4Gi default.
Dependency drift — a --no-cache local build resolves byte-identical package versions and starts in ~2s.
Application code — reproduced with a stock agents-cli create project (adk + agent_runtime) in a clean directory.

One reporting note that may explain why some of these look like silent hangs:
the cloudaudit.googleapis.com/activity record's status is sometimes an empty {} even when the deploy failed. The real error only appears on the operation resource:

GET https://{LOCATION}-aiplatform.googleapis.com/v1beta1/{ENGINE}/operations/{OP_ID}
-> { "done": true, "error": { "code": 3, "message": "... failed to start ..." } }

Environment: google-adk 2.6.1, a2a-sdk 0.3.26 / 1.1.2 (both tried), agents-cli 1.2.1, Python 3.12, deploying to Agent Runtime.

So the issue that i was experiencing was actually an infrastructure problem

After opening a support case, i received this from Google:

The deployment failures in us-west1 are caused by an active known issue we have identified, and we’re currently working on on our end. Since this happens at the infrastructure orchestration layer before the container starts up, no application/runtime logs are generated.

Deploy the production workloads temporarily to us-east1 or another healthy alternative region.

If your workloads need to remain in the us-west1 region, we highly recommend reducing the CPU and memory requests [1] in your deployment specifications. Lowering your resource footprint significantly decreases the chances of scheduling delays or rejections during times of high regional demand.

As of 10 August 2026, this issue has been resolved

@rdamus

Closing this out. Two separate things got tangled into one thread here. The regional "failed to start" errors (europe-west2, then us-west1) were a Vertex AI Agent Runtime infra issue confirmed fixed by Google as of Aug 10 per @rdamus's update above, nothing ADK-side to do there. Separately the old templated import that #4243 was patching (broke on hyphenated names / agents under app/) has since been replaced by a rewritten deploy path that validates the app name and imports the agent dynamically via importlib so that failure mode is gone on current versions too. #4243 is superseded rather than pending not something that still needs merging. @Alameyo if you're still hitting this on 2.8.0+ and we'll reopen.

Navigation Menu

Repository navigation

ADK fails to deploy with 1.18.0 and above #4237

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

ADK fails to deploy with 1.18.0 and above #4237

Description

Activity

smwitkowski commented on Jan 22, 2026

llalitkumarrr commented on Feb 2, 2026

francamacdowell commented on Mar 2, 2026

llalitkumarrr commented on Mar 5, 2026

adk-bot commented on Mar 12, 2026

Alameyo commented on Mar 12, 2026

llalitkumarrr commented on Mar 16, 2026

Alameyo commented on Mar 21, 2026

llalitkumarrr commented on Mar 26, 2026

rdamus commented on Aug 5, 2026

rdamus commented on Aug 10, 2026

surajksharma07 commented on Sep 8, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions