VertexAiSessionService does not support Resumability

[lucasbarzotto-axonify]

🔴 Required Information

Describe the Bug:

Resumability doesn't work when using VertexAiSessionService. I believe this is related to the TODO in vertex_ai_session_service.pyline 270:

Since the agent_state and end_of_agent fields are not sent to VertexAiSessionService when appending an Event, it seems like "empty" Events are created (events that should include agent_state or end_of_agent information):

On top of Resumability not working, the following unexpected behavior was observed:

• As a consequence of "empty" events being appended, the _find_agent_to_run method from runners.py doesn't work as expected. The last event appended is always from the root_agent, so the selected agent to run is always the root_agent.
• To reproduce this, use VertexAiSessionService with Resumability enabled and create a root_agent with at least 2 sub_agents. You can see that the conversation always go back to the root_agent, instead of continuing on the sub_agent as expected, for example:

Steps to Reproduce:

1. Pass an agent engine uri for the session service, so VertexAiSessionService is used
2. Add ResumabilityConfig with is_resumable=True set
3. Create an agent with at least 2 sub agents

Example:

import logging

from google.adk.agents import Agent
from google.adk.apps import App, ResumabilityConfig

logger = logging.getLogger(__name__)

travel_agent = Agent(
    name="travel_agent",
    model="gemini-2.5-flash",
    description="The travel agent.",
    instruction="You are a helpful agent who can help the user with their travel plans.",
    tools=[],
)


doctor_agent = Agent(
    name="doctor_agent",
    model="gemini-2.5-flash",
    description="The doctor agent.",
    instruction="You are a helpful agent who can help the user with their health questions.",
    tools=[],
)

root_agent = Agent(
    name="helpful_agent",
    model="gemini-2.5-flash",
    description="The orchestrator agent.",
    instruction="You are a helpful agent who can greet the user and transfer to the appropriate sub-agent.",
    tools=[],
    sub_agents=[travel_agent, doctor_agent],
)

app = App(
    name="helpful_agent",
    root_agent=root_agent,
    resumability_config=ResumabilityConfig(  # Comment this out so it works
        is_resumable=True,
    ),
)

Expected Behavior:

VertexAiSessionService should support Resumability, or at least it should be mentioned in the documentation that this is not supported yet.

Observed Behavior:

VertexAiSessionService doesn't send the agent_state and end_of_agent fields when appending an Event. As a consequence, Resumability doesn't work and _find_agent_to_run doesn't work as it should.

Environment Details:

• ADK Library Version (pip show google-adk): 1.23
• Desktop OS: macOs
• Python Version (python -V): 3.13.11

Model Information:

• Are you using LiteLLM: No
• Which model is being used: gemini-2.5-flash

---

🟡 Optional Information

Providing this information greatly speeds up the resolution process.

Regression:
Did this work in a previous version of ADK? If so, which one? I don't think so. I think it never worked before.

How often has this issue occurred?:

• Always (100%)

[lucasbarzotto-axonify]

Do you have any estimates of when VertexAiSessionService will support ResumabilityConfig?

[adk-bot]

Notification: The author has updated the issue description. Maintainers, please review.

[llalitkumarrr]

Hello @lucasbarzotto-axonify ,

Thank you for providing the code. We tested the issue using google-adk==1.24.1 and could not reproduce the problem.

We used the exact code you sent with the Vertex AI Session Service with ADK web. All health related inquiries were correctly routed to the doctor_agent. We then switched to travel related questions and the system correctly transferred those to the travel_agent as expected.

[lucasbarzotto-axonify]

Hi @llalitkumarrr, thanks for your response!

From what I understand, the behavior shown in your screenshot is not the intended behavior. In your screenshot for every new message from the user, the root_agent is executed again, and it needs to transfer to one of the sub_agents.

From what I understand, the correct behavior would be to run the current sub_agent automatically, without having the root_agent transfer to it first.

This "automatic run" logic exists in the function _find_agent_to_run in runners.py

I believe if you test this exact setup with Resumability turned off, you should be able to see that the transfer_to_agent calls doesn't happen for each new message, but the last sub_agent automatically continues the conversation.

For example:

Note that the transfer_to_agent doesn't happen for each new message. The doctor agent is executed automatically.

[llalitkumarrr]

Hello @lucasbarzotto-axonify ,

I understand your point, we are using the agent as a sub-agent rather than an AgentTool so it should retain the full control.
This is reproducible with adk run as well.

You mentioned you might have identified the cause in the code; could you share more details? We would value your contribution if you are willing to help us fix this by submitting a PR.

[llalitkumarrr]

Hello @wyf7107 ,

Could you please have a look at this?

[lucasbarzotto-axonify]

I believe the issue is related to the following TODO comment in vertex_ai_session_service.py line 270:

Note how it mentions that agent_state is not supported by the API. agent_state is directly related to the resumability feature.

Would you be able to let us know whether ADK plans to support resumability with VertexAiSessionService?

Thanks again!