"ADK Bug: before_model_callback causes AttributeError on Agent Engine"

[Saoussen-CH]

ADK Bug Report: before_model_callback causes AttributeError on Agent Engine

Bug Summary

When deploying an ADK agent with before_model_callback to Vertex AI Agent Engine, the agent fails with:

AttributeError: 'LlmRequest' object has no attribute 'content'

This error occurs in ADK's internal _nl_planning.py:79 file, preventing the agent from processing any requests.

Impact

• Severity: Critical - Agent cannot process any requests
• Scope: Affects all agents using before_model_callback on Agent Engine
• Workaround: None - must disable before_model_callback entirely

Environment

• ADK Version: google-cloud-aiplatform[adk,agent_engines]>=1.112
• Python Version: 3.11
• Deployment Target: Vertex AI Agent Engine (not local runner)
• Project: project-ddc15d84-7238-4571-a39
• Location: us-central1

Steps to Reproduce

1. Create an agent with before_model_callback:

from google.adk.agents import Agent

async def my_callback(callback_context, llm_request):
    """Simple callback that modifies system instruction"""
    print("Callback executed")
    # Any modification here triggers the bug
    return llm_request

agent = Agent(
    name="test_agent",
    model="gemini-2.0-flash-exp",
    instruction="Test agent",
    tools=[],
    before_model_callback=my_callback  # THIS TRIGGERS THE BUG
)

2. Deploy to Agent Engine:

import vertexai
from vertexai import agent_engines

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

remote_app = agent_engines.create(
    agent=agent,
    requirements=["google-cloud-aiplatform[adk,agent_engines]>=1.112"],
    staging_bucket="gs://your-bucket"
)

3. Query the deployed agent:

response = remote_app.query(user_id="test", message="Hello")

4. Observe the error in logs:

ERROR: AttributeError: 'LlmRequest' object has no attribute 'content'

Full Error Traceback

Traceback (most recent call last):
  File ".../vertexai/agent_engines/templates/adk.py", line 971, in async_stream_query
    async for event in events_async:
  File ".../google/adk/runners.py", line 454, in run_async
    async for event in agen:
  File ".../google/adk/agents/llm_agent.py", line 460, in _run_async_impl
    async for event in agen:
  File ".../google/adk/flows/llm_flows/base_llm_flow.py", line 346, in run_async
    async for event in agen:
  File ".../google/adk/flows/llm_flows/base_llm_flow.py", line 433, in _run_one_step_async
    async for event in agen:
  File ".../google/adk/flows/llm_flows/base_llm_flow.py", line 510, in _postprocess_async
    async for event in agen:
  File ".../google/adk/flows/llm_flows/base_llm_flow.py", line 657, in _postprocess_run_processors_async
    async for event in agen:
  File ".../google/adk/flows/llm_flows/_nl_planning.py", line 79, in run_async
    or not llm_response.content
           ^^^^^^^^^^^^^^^^^^^^
  File ".../pydantic/main.py", line 1026, in __getattr__
    raise AttributeError(f'{type(self).__name__!r} object has no attribute {item!r}')
AttributeError: 'LlmRequest' object has no attribute 'content'

Root Cause Analysis

In _nl_planning.py:79, the code tries to access llm_response.content:

or not llm_response.content

However, after before_model_callback executes, the variable is actually an LlmRequest object (not LlmResponse), which doesn't have a content attribute.

This appears to be a type confusion bug in ADK's internal flow processing.

Expected Behavior

• The before_model_callback should execute successfully
• The agent should continue processing and return a response
• Memory loading, prompt modification, and other callback use cases should work

Actual Behavior

• The callback executes successfully
• ADK's internal _nl_planning.py immediately crashes
• Agent returns error code 498 with the AttributeError message
• No response is generated

Use Case

We're trying to implement Memory Bank integration with a custom callback (since PreloadMemoryTool has its own bugs). This is the recommended approach per Google's documentation:

• https://github.com/GoogleCloudPlatform/generative-ai/blob/main/agents/agent_engine/memory_bank/get_started_with_memory_bank_on_adk.ipynb

Workaround

None - The only option is to completely disable before_model_callback, which removes the ability to:

• Load memories before each turn
• Modify prompts dynamically
• Implement custom preprocessing logic

Additional Context

• The bug only occurs on Agent Engine (not local runner)
• The bug occurs regardless of what the callback does (even an empty callback triggers it)
• Multiple deployments confirmed the same error
• Agent Engine resource IDs where bug was observed:
  • projects/773461168680/locations/us-central1/reasoningEngines/1794943936251297792
  • projects/773461168680/locations/us-central1/reasoningEngines/3116750426884538368
  • projects/773461168680/locations/us-central1/reasoningEngines/6440406951883964416

Logs

Cloud Logging query to see the error:

resource.type=aiplatform.googleapis.com/ReasoningEngine
AND resource.labels.reasoning_engine_id="6440406951883964416"
AND severity=ERROR

Request

Please fix the type confusion in _nl_planning.py:79 to properly handle LlmRequest objects returned from before_model_callback.

[ankursharmas]

Please refer to the guidance on how callbacks should be implemented: https://google.github.io/adk-docs/callbacks/types-of-callbacks/#after-agent-callback

The call back method should either return None to ensure that the processing continues or the call back should return an object of type LLM Response to essentially short-circuit the call the model.

If your intent was to perform some kind of mutations to LLMRequest object, then you are free to do so. Your callback method gets the same object that will be passed down to the model.