Inline source deployment incompatible with agent_server_mode: EXPERIMENTAL

[ASRagab]

Summary

When deploying an agent to Vertex AI Agent Engine using inline source deployment (source_packages, entrypoint_module, entrypoint_object) with agent_server_mode: EXPERIMENTAL, the deployment fails at runtime. The EXPERIMENTAL server runtime attempts to load the agent from a pickle file instead of using the inline source code.

Environment

• google-cloud-aiplatform: 1.135.0
• google-adk: latest
• Python: 3.11/3.12
• Region: us-central1

Steps to Reproduce

1. Create an ADK agent with AdkApp
2. Deploy using the GenAI SDK with inline source configuration:

client.agent_engines.create(config={
    "source_packages": ["my_agent_package", "requirements.txt"],
    "entrypoint_module": "runtime_wrapper",
    "entrypoint_object": "adk_app",
    "class_methods": [...],
    "agent_server_mode": vertexai_types.AgentServerMode.EXPERIMENTAL,
})

3. Deployment completes but the agent fails to start

Expected Behavior

The agent should start successfully using the inline source code, similar to how it works without agent_server_mode: EXPERIMENTAL.

Actual Behavior

The agent fails with the following error in Cloud Logging:

FileNotFoundError: [Errno 2] No such file or directory: 'user_code/code.pkl'

Stack trace shows the EXPERIMENTAL server runtime expects a pickle file:

File "/code/server.py", line 61, in serve
    servicer = service_builder.construct_service(pickle_file)
File "/code/service_builder.py", line 349, in construct_service
    obj = utils.get_object(python_file_name)

Analysis

The issue appears to be that:

1. Without EXPERIMENTAL mode: The server runtime correctly uses inline source (entrypoint_module + entrypoint_object)
2. With EXPERIMENTAL mode: A different server runtime is used that only supports pickle-based deployment

Removing agent_server_mode: EXPERIMENTAL from the config allows the same agent to deploy and run successfully with inline source.

Questions

1. Is this a known limitation of EXPERIMENTAL mode?
2. Is there a timeline for inline source support in EXPERIMENTAL mode?
3. Is pickle-based deployment currently the only supported method for bidirectional streaming agents?

Workaround

Currently, the only workaround is to use pickle-based deployment (with staging_bucket and agent= parameters) for agents requiring EXPERIMENTAL mode, which introduces complexity around GCS bucket management and pickle serialization compatibility.

References

• Inline Source Deployment Discussion
• Bidirectional Streaming Documentation