[Live/bidi] transfer_to_agent crashes with 1011 on Agent Engine — sub_agents delegation not supported on bidi_stream_query

[smwitkowski]

Required Information

Describe the Bug

When using bidi_stream_query() on Agent Engine, any agent that registers sub-agents via sub_agents=[...] crashes with a 1011 (internal error) WebSocket close when the model calls the built-in transfer_to_agent function. The sub-agent never executes. The session is terminated immediately.

Crash rate: 100% when transfer_to_agent fires (10/10 trials in isolated reproduction).

This is specific to the bidi/live path on Agent Engine. The same agent with the same sub-agents works correctly via Runner.run_async locally (10/10 PASS — sub-agent delegation succeeds, response is produced).

Workaround: Wrapping the sub-agent in AgentTool(agent) instead of registering it via sub_agents=[] avoids the crash. AgentTool registers a regular function call (not transfer_to_agent) that works on the bidi path.

Steps to Reproduce

1. Create subagent_crash/agent.py:

from google.adk.agents import Agent
from google.adk.models import Gemini
from google.genai.types import GenerateContentConfig

faq_agent = Agent(
 name="faq_agent",
 description="Answers all user questions",
 model=Gemini(model="gemini-2.5-flash"),
 instruction="Answer the user's question concisely in 1-2 sentences.",
 tools=[],
)

root_agent = Agent(
 model=Gemini(model="gemini-live-2.5-flash-native-audio"),
 name="root_agent",
 instruction=(
 "You are a routing agent. You MUST delegate ALL user messages to faq_agent "
 "using the transfer_to_agent function. "
 "NEVER answer any question yourself. NEVER respond directly. "
 "For EVERY single user message, immediately call transfer_to_agent with "
 "agent_name='faq_agent'. No exceptions."
 ),
 sub_agents=[faq_agent],
 tools=[],
 generate_content_config=GenerateContentConfig(temperature=0.0),
)

2. Create subagent_crash/agent_engine_app.py:

from vertexai.preview.reasoning_engines.templates.adk import AdkApp
from subagent_crash.agent import root_agent

class AgentEngineApp(AdkApp):
 def register_operations(self):
 operations = super().register_operations()
 operations["bidi_stream"] = ["bidi_stream_query"]
 return operations

agent_engine = AgentEngineApp(agent=root_agent)

3. Deploy to Agent Engine, connect via bidi, and send any message:

import asyncio
import vertexai

async def test():
 client = vertexai.Client(project="YOUR_PROJECT", location="us-central1")
 async with client.aio.live.agent_engines.connect(
 agent_engine="projects/.../reasoningEngines/..."
 ) as session:
 await session.send(query_input={"user_id": "test"})
 await session.send(query_input={
 "content": {"role": "user", "parts": [{"text": "Hello"}]}
 })
 while True:
 event = await asyncio.wait_for(session.receive(), timeout=15)
 print(event)

asyncio.run(test())

Expected Behavior

The model calls transfer_to_agent(agent_name="faq_agent"), ADK transfers control to faq_agent, the sub-agent responds, and the response is streamed back via the bidi session.

Actual Behavior

Event 1: function_call: transfer_to_agent(agent_name="faq_agent")
Event 2: function_response: transfer_to_agent → result: null
 → ConnectionClosedError: received 1011 (internal error)
 → "Reasoning Engine Execution failed"

The WebSocket closes immediately after the transfer_to_agent function response. Only 4 events are produced. No content is delivered to the user.

Key Distinction: sub_agents vs AgentTool

This crash is specific to the sub_agents=[] parameter, which registers the built-in transfer_to_agent function. Wrapping the same agent in AgentTool() and placing it in tools=[] registers a regular function call and does not crash:

# CRASHES on bidi — uses transfer_to_agent internally
root_agent = Agent(
 sub_agents=[faq_agent],
 ...
)

# WORKS on bidi — uses regular function call
from google.adk.tools import AgentTool
root_agent = Agent(
 tools=[AgentTool(faq_agent)],
 ...
)

Both patterns work identically via Runner.run_async locally. The crash is only on the Agent Engine bidi path.

Local Baseline

Same agent tested locally via Runner.run_async with gemini-2.5-flash: 10/10 PASS. transfer_to_agent is called, the sub-agent runs, a response is produced. The crash is specific to Agent Engine's bidi_stream_query.

Environment

• google-adk==1.28.1
• google-cloud-aiplatform==1.145.0
• Model: gemini-live-2.5-flash-native-audio (root), gemini-2.5-flash (sub-agent)
• Path: bidi_stream_query on Vertex AI Agent Engine
• Python 3.10
• Region: us-central1

Related Issues

• [Live] run_live(): Fire-and-forget tool calls cause duplicate model responses via orphaned function response reinjection #4902 — Duplicate responses from orphaned tool re-entry (fixed in 1.28.1, different mechanism)
• [Live/bidi_stream_query] Session state silently dropped — first queue message 'state' field never read, async_create_session called without state= #5113 — Session state loss in bidi_stream_query (fix merged, pending release)

[smwitkowski]

Here's the full deploy script to reproduce on any GCP project with Agent Engine:

# deploy.py
import vertexai
from vertexai._genai.types import AgentEngineConfig, AgentServerMode
from vertexai._genai import _agent_engines_utils
from subagent_crash.agent_engine_app import agent_engine

project = "YOUR_PROJECT"
location = "us-central1"

client = vertexai.Client(project=project, location=location)
vertexai.init(project=project, location=location)

registered_ops = _agent_engines_utils._get_registered_operations(agent=agent_engine)
class_methods_spec = _agent_engines_utils._generate_class_methods_spec_or_raise(
 agent=agent_engine, operations=registered_ops,
)
class_methods = [_agent_engines_utils._to_dict(m) for m in class_methods_spec]

config = AgentEngineConfig(
 display_name="subagent-crash-repro",
 description="Minimal repro: transfer_to_agent crash on bidi",
 extra_packages=["./subagent_crash"],
 requirements="requirements.txt",
 staging_bucket=f"gs://{project}-agent-engine",
 agent_server_mode=AgentServerMode.EXPERIMENTAL,
 class_methods=class_methods,
 env_vars={
 "GOOGLE_CLOUD_PROJECT_ID": project,
 "GOOGLE_GENAI_USE_VERTEXAI": "True",
 },
)

remote = client.agent_engines.create(agent=agent_engine, config=config)
print(f"Deployed: {remote.api_resource.name}")

With requirements.txt:

google-adk>=1.28.1,<2.0.0
google-cloud-aiplatform[agent-engines]>=1.132.0
google-genai>=1.0.0,<2.0.0
cloudpickle>=3.0.0
pydantic>=2.0.0

Re: the model question — not a typo. Root agent uses gemini-live-2.5-flash-native-audio (voice/bidi session) and the sub-agent uses gemini-2.5-flash (text processing). This matches our production setup. The crash fires on the transfer_to_agent call before the sub-agent's model is ever invoked.

One note: a teammate tested a similar setup on a different AE project and did not see the crash, so it may be environment-specific. Let me know if you're able to reproduce, and if not I can share our exact project/agent IDs.

[wuliang229]

The issue is from model mis-use. The sub-agent should also use a Live model.