No logging output and unexpected result during Agent Engine deployment

[cychoi11]

The execution of the deployment command (client.agent_engines.update) pauses without printing any operational logs. Upon completion of the synchronous call, it returns an error and prints the specific, unhelpful message:

"Failed to update Agent Engine: {code: 10 ...}" along with a generic request to check the documentation.

error message

RuntimeError: Failed to update Agent Engine : { 'code' : 'message' : 'please refer to our documentation (https://cloud.google.com/vertex-ai/generative-ai/docs/agent-engine/troubleshootiong/deploy) for checking logs and other troubleshooting tips.

Crucially, after this failure, I checked the corresponding Cloud Logging (Vertex AI Agent Engine/Operation logs) for the relevant project and region, but there are NO error logs or warning messages that correlate with this specific failure time or Agent Engine ID.

Expected Behavior
I expect the client.agent_engines.update() call to successfully deploy the Agent Engine and return a fully updated remote_a2a_agent object.

If the deployment fails, I expect one of the following to happen:

Local Logging: The execution should print detailed operational logs (HTTP requests and polling status) to the Jupyter output stream during the process to help diagnose local serialization or environment issues.

Remote Logging: A clear and actionable error message must be visible in the corresponding GCP Cloud Logging (Vertex AI Agent Engine/Operation logs) to facilitate remote troubleshooting of build or infrastructure issues.

This indicates a failure during the remote Agent Engine update process, but the absence of logs prevents any local diagnosis.

def get_data(purpose: str, user_input: str, tool_context: ToolContext) -> Dict[str, Any]:
    # 세션 상태에서 검증 완료 여부 가져오기 (기본값 False)
    
    data = pd.read_excel("./user_data1.xlsx")
    data_dict = data.to_dict(orient='records')

    return {
        "data": data_dict
    }

review_agent = LlmAgent(
    name="review_agent",
    model="gemini-2.5-flash",
    # Change 3: Improved instruction
    instruction="""당신은 사용자의 질문에 따라 데이터를 가져와서 보고하는 에이전트입니다.""",
    description="당신은 사용자의 질문에 따라 데이터를 가져와서 보고하는 에이전트입니다.",
    tools=[get_data],
    output_key="review"
)

varification = LlmAgent(
    name="varification",
    model="gemini-2.5-flash",
    # Change 3: Improved instruction
    instruction="""당신은 review 에이전트의 결과 {review}를 확인하고 이것을 통해 사용자가 보기좋게 보고서를 작성하는 역할을 합니다.""",
    description="당신은 review 에이전트의 결과를 확인하고 이것을 통해 사용자가 보기좋게 보고서를 작성하는 역할을 합니다.",
)




seq_agent = SequentialAgent(
    name="seq_agent",
    sub_agents=[review_agent,varification],
    description="차례대로 실행되는 에이전트의 흐름"
)

report = AgentSkill(
    id="report",
    name="report",
    description="지금까지의 동작 내용을 사용자에게 설명하는 agent 입니다.",
    tags=["report"],
)

sequential_agent_card = create_agent_card(
    agent_name='code_pipeline_agent',
    description='필요에 따라 작업을 분배하고 작업에 맞게 에이전트를 호출하여 작업을 수행하세요',
    skills=[
    report
    ]
)

from google.adk.auth.credential_service.in_memory_credential_service import InMemoryCredentialService
async def create_runner() -> Runner:
    """Create a runner for the agent."""
    return Runner(
        app_name=seq_agent.name or "adk_agent",
        agent=seq_agent,
        # Use minimal services - in a real implementation these could be configured
        artifact_service=GcsArtifactService(bucket_name=BUCKET_NAME),
        session_service=VertexAiSessionService(project= PROJECT_ID, location= LOCATION, agent_engine_id= AGNET_ENGINE_ID),
        memory_service=InMemoryMemoryService(),
        credential_service=InMemoryCredentialService(),
    )
task_store = InMemoryTaskStore()

agent_executor = a2a_agent_executor.A2aAgentExecutor(
    runner=create_runner
)

request_handler = DefaultRequestHandler(
    agent_executor=agent_executor, task_store=task_store
)

sequential_a2a = A2aAgent(
    agent_card= sequential_agent_card,
    request_handler_builder=request_handler
)

sequential_a2a.set_up()

The deployment is executed in an update fashion to a pre-existing Agent Engine.

remote_a2a_agent = client.agent_engines.update(
    # The actual agent to deploy
    name=f"projects/{PROJECT_ID}/locations/{LOCATION}/reasoningEngines/{AGENT_ENGINE_ID}",
    agent=sequential_a2a,
    config={
        "display_name": "seq_executor_agent",
        "requirements": [
            "google-cloud-aiplatform[agent_engines,adk]>=1.117.0",
            "a2a-sdk >= 0.3.7",
            "cloudpickle >= 3.1.1",
            "pydantic >= 2.11.9",
            "pandas"
        ],
        # Staging bucket
        "staging_bucket":  "gs://{BUCKET_NAME}",
    },
)

Please advise on how to enforce verbose logging for the agent_engines.update method and investigate this silent remote failure. Thank you for your support.

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: The deployment is likely failing because your get_data tool attempts to read a local file (./user_data1.xlsx), which is not accessible in the remote Agent Engine environment.

---

Hello @cychoi11,

Thank you for the detailed report. This issue, where the deployment fails silently without clear logs, can be frustrating.

Based on the code you've provided, the root cause of the error is almost certainly the local file dependency in your get_data tool.

Explanation

When you call client.agent_engines.update, the ADK packages your agent's code and its dependencies and sends them to a managed environment in Google Cloud to run. [1, 2] This remote environment is completely separate from your local machine (or Jupyter notebook environment) and does not have access to your local filesystem.

The get_data tool contains the line:
data = pd.read_excel("./user_data1.xlsx")

The deployment process fails when it tries to set up your agent in the cloud because it cannot find the file ./user_data1.xlsx. This failure happens during the backend build process, which is why you are not seeing any specific error messages in your local output or in the standard Cloud Logging channels.

Recommended Solution

To fix this, you should store your data file in a location that the remote agent can access, such as a Google Cloud Storage (GCS) bucket.

1. Upload the data file: Upload user_data1.xlsx to a GCS bucket that your project's service account can access.

2. Update your tool code: Modify the get_data tool to read the file from GCS. You can do this easily with pandas if you also install the gcsfs library.

   First, add gcsfs and fsspec to your deployment requirements:

   remote_a2a_agent = client.agent_engines.update(
       name=f"projects/{PROJECT_ID}/locations/{LOCATION}/reasoningEngines/{AGENT_ENGINE_ID}",
       agent=sequential_a2a,
       config={
           "display_name": "seq_executor_agent",
           "requirements": [
               "google-cloud-aiplatform[agent_engines,adk]>=1.117.0",
               "a2a-sdk >= 0.3.7",
               "cloudpickle >= 3.1.1",
               "pydantic >= 2.11.9",
               "pandas",
               "gcsfs",         # <-- Add this
               "fsspec",        # <-- Add this
               "openpyxl"       # <-- Add for Excel support in pandas
           ],
           "staging_bucket": f"gs://{BUCKET_NAME}",
       },
   )

   Then, update your get_data function to point to the GCS path:

   def get_data(purpose: str, user_input: str, tool_context: ToolContext) -> Dict[str, Any]:
       # Path to your file in GCS
       gcs_path = "gs://YOUR_BUCKET_NAME/path/to/user_data1.xlsx"
       
       # Read the excel file directly from GCS
       data = pd.read_excel(gcs_path)
       data_dict = data.to_dict(orient='records')
   
       return {
           "data": data_dict
       }

Regarding Verbose Logging

You are right to expect better logging. The lack of detailed error feedback during the remote build process is a known challenge. As you've discovered, client-side logging (e.g., via logging.basicConfig) wouldn't help in this scenario because the error is not happening on the client. [3] The generic message is all that the server returns after the backend failure. We are continuously working to improve the observability of the deployment process.

By resolving the file path issue, your deployment should succeed. Please let us know if you encounter any further issues.

---

[1] https://google.github.io/adk-docs/deploy/agent-engine/
[2] https://github.com/google/adk-python/blob/main/contributing/adk_project_overview_and_architecture.md
[3] https://google.github.io/adk-docs/observability/logging/

[cychoi11]

Hello Team,

Thank you for your previous guidance on the deployment failure (code: 10), which we initially suspected was due to a local file dependency.

Following your advice, I have completely restructured the get_data tool to remove the local file dependency (pd.read_excel) and now it returns a statically defined Python dictionary.

Despite removing this file access issue, the deployment still fails silently when executing client.agent_engines.update(), returning a generic RuntimeError (e.g., {code : 10...}) and no logs in GCP Cloud Logging.

1. Current State

The problem persists even after the primary suspected cause was eliminated. The Agent Engine cannot be deployed, and we have no visibility into the cause.

Environment

• Platform: Google Agent Engine Workbench
• Python Version: 3.10.18

Modified Tool Code

The get_data function is now clean, returning only pure Python types:

def get_data(purpose: str, user_input: str, tool_context: ToolContext) -> Dict[str, Any]:
    # 세션 상태에서 검증 완료 여부 가져오기 (기본값 False)
    
    data_dictionary = {
    "번호": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
    "이름": ["김철수", "이영희", "박민지", "최현우", "정**", "송**", "한**", "탈퇴유저_T008", "탈퇴유저_T009", "탈퇴유저_T010"],
    "주민등록번호": [
        "900101-1234567", "950515-2765432", "881224-1505050", "750820-1919191", 
        "990310-2******", "820722-1******", "021105-4******", None, None, None
    ],
    "전화번호": [
        "010-1234-5678", "010-9876-5432", "010-3333-7777", "010-2222-8888",
        "010-4444-****", "010-5555-****", "010-6666-****", None, None, None
    ],
    "처리 범주 (작업내용)": [
        "작업 X (전체 공개)", "작업 X (전체 공개)", "작업 X (전체 공개)", "작업 X (전체 공개)",
        "마스킹 처리", "마스킹 처리", "마스킹 처리", "탈퇴/삭제 처리", "탈퇴/삭제 처리", "탈퇴/삭제 처리"
    ]
    }
    return data_dictionary

2. ❓ Assistance Required

Given that the failure persists and no useful logs are generated locally or remotely, I need assistance with the following:

1. GCP Logging Visibility: Are there any non-standard methods or configurations (e.g., specific environment variables, different log sinks, or internal tools) that can force the Agent Engine's remote build process to emit debug logs to Cloud Logging when it encounters an early failure?
2. Deployment Configuration Review: Based on the common reasons for silent Agent Engine failures (other than file access), what are the most likely issues related to the deployment configuration (config parameters) or IAM permissions that I should check next?

Your help in achieving visibility into this deployment failure is critical. Thank you.