ParallelAgent works with LLMAgent but not with custom agents — is additional setup needed?

[shewalesantosh06-sant]

I'm working with the ADK Agents and have encountered an issue when using ParallelAgent with custom agents.

Context:
I have created 3 custom agents (not using LLMAgent) and grouped them under a ParallelAgent.

When running, I notice that these agents are executed sequentially(one after the other), not in parallel as expected.

However, if I replace the custom agents with LLMAgent instances, the parallel execution works as intended — the LLM calls are dispatched concurrently.

Questions:
Why does ParallelAgent support true parallel execution with LLMAgent but not with custom agents?

Is there any additional configuration or specific interface/behavior required for custom agents to be compatible with ParallelAgent and actually run in parallel?

parallel agent code :

parallel_extraction_agent = ParallelAgent(
name="parallel_extraction_agent",
description="This agent orchestrates multiple sub agents to extract specific information from the medical record",
sub_agents=[accommodations_extraction_agent, info_agent, synopsis_agent],
)

[dexhunter]

ParallelAgent also supports custom BaseAgent subclasses. It runs their async event streams concurrently; it does not automatically move synchronous work into separate threads.

Check the implementation of each custom agent's _run_async_impl. A synchronous HTTP request, time.sleep(), or a long computation inside that method blocks the shared event loop, even though the method is declared async.

For blocking I/O, use an async client and await it, or move the existing synchronous operation to a thread:

import asyncio

# Inside your custom agent's async _run_async_impl:
result = await asyncio.to_thread(blocking_extract, document)
# Yield your normal ADK Event containing the result here.

asyncio.to_thread is primarily for blocking I/O; CPU-bound Python work generally needs a process or an external worker to run in parallel. Also ensure each agent has independent input/output state keys if it writes shared session state.

There is no extra ParallelAgent setting to enable custom-agent concurrency. Its implementation schedules the sub-agents' run_async() streams as asyncio tasks. The wrapper shown is sufficient; the custom _run_async_impl code is needed to identify the particular blocking operation.