Context Compaction question

[mmarzol-hatchworks]

Hi everyone,

Quick question. I'm trying to use the Context Compaction feature (google-adk 1.17.0), and defined the agents in the following way:

# 1. Define the agent
root_agent = SequentialAgent(
    name="root", sub_agents=[agent_parallel_main, agent_synthesizer]
)

# 2. Create an App instance, passing in Events Compaction config for context compaction.
main_app = App(
    name='myApp',
    root_agent=root_agent,
    events_compaction_config=EventsCompactionConfig(
        compaction_interval=3,
        overlap_size=1
    ),
)

# 3. Export the app for use in the main.py file. The deployment tools will discover this.
app = main_app

However, when checking the Cloud Run logs for the container running the ADK, I can see the follwoing entries:

Info: future: <Task finished name='Task-2309' coro=<_run_compaction_for_sliding_window() done, defined at /usr/local/lib/python3.13/site-packages/google/adk/apps/compaction.py:27> exception=AttributeError("'SequentialAgent' object has no attribute 'canonical_model'")>

Error: File "/usr/local/lib/python3.13/site-packages/google/adk/apps/compaction.py", line 186, in _run_compaction_for_sliding_window
llm=app.root_agent.canonical_model
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.13/site-packages/pydantic/main.py", line 1026, in getattr
raise AttributeError(f'{type(self).name!r} object has no attribute {item!r}')
AttributeError: 'SequentialAgent' object has no attribute 'canonical_model'

Is there any errors on how I am defining the App with EventsCompactionConfig?

Thank you!

[adk-bot]

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

TLDR: The AttributeError occurs because your root_agent is a SequentialAgent, which does not have a language model (canonical_model) attached to it. The context compaction feature requires a model to summarize events. To fix this, you need to explicitly provide a summarizer with a designated model within your EventsCompactionConfig.

---

Hi @mmarzol-hatchworks,

Thank you for your question. I can see you're running into an issue where the context compaction process is unable to find a language model.

Why the Error Happens

The context compaction feature works by summarizing past events into a condensed form. This summarization is an LLM-powered operation. By default, the system tries to get the required LLM from your root_agent by accessing app.root_agent.canonical_model.

In your setup, the root_agent is a SequentialAgent. SequentialAgent is a workflow agent designed to orchestrate and route tasks between its sub_agents; it does not have its own canonical_model. This is why the system raises an AttributeError.

How to Fix It

To resolve this, you can explicitly provide a model for the compaction process. This is done by creating an LlmEventSummarizer (or a custom summarizer) and passing it to the EventsCompactionConfig. This tells the App to use a specific model for summarization tasks, rather than looking for one on the root_agent.

Here is your code, modified with the solution:

from google.adk.apps.app import App, EventsCompactionConfig
from google.adk.apps.llm_event_summarizer import LlmEventSummarizer
from google.adk.models import Gemini
from google.adk.agents import SequentialAgent #, other agent types

# Assuming agent_parallel_main and agent_synthesizer are defined
# agent_parallel_main = ...
# agent_synthesizer = ...

# 1. Define the agent
root_agent = SequentialAgent(
    name="root", sub_agents=[agent_parallel_main, agent_synthesizer]
)

# 2. Create a summarizer with a specific model for compaction.
# This model will be used to summarize the events.
compaction_summarizer = LlmEventSummarizer(
    llm=Gemini(model="gemini-1.5-flash") # Or another model of your choice
)

# 3. Create an App instance, passing the custom summarizer in the config.
main_app = App(
    name='myApp',
    root_agent=root_agent,
    events_compaction_config=EventsCompactionConfig(
        compaction_interval=3,
        overlap_size=1,
        summarizer=compaction_summarizer  # Pass the summarizer here
    ),
)

# 4. Export the app for use in the main.py file.
app = main_app

By providing a summarizer, the compaction process will use the Gemini model instance you defined for summarization, and it will no longer need to access the canonical_model of the root_agent.

You can find more details on this in the official documentation. [1]

Hope this helps!

[1] https://google.github.io/adk-docs/context/compaction/