Is it possible to call multiple instances of one subagent in parallel?

[consilience-dev]

I have a tool that gives me a list of outputs, say 1 to N outputs. These outputs then need to be processed, but I'd like to send each one to a separate agent. Specifically, I'd like to only define one sub_agent that would be able to handle each of these outputs.

After, I'd like to merge the results with some merger subagent. Is this a valid use case? Is there some way to do this?

[zbigniev]

I believe, the example here does exactly what you are looking for https://google.github.io/adk-docs/agents/workflow-agents/parallel-agents/

[consilience-dev]

Hi @zbigniev , thanks for the reply. In my use case I have one sub_agent, let's call it "comparer_agent".

I have some root_agent which does a tool call, and gets back N results. I want to send each of these results to an instance of "comparer_agent". Specifically, in my parallel call, I'd like between 1 to N instances of my comparer_agent called in parallel.

In the example from the ADK docs that you provided, there is a hard number of research_agents; 3 specifically.

I'm wondering if I can have an arbitrary number of parallel agent executions while only specifying one comparer_agent class

[zbigniev]

Maybe, I misunderstood what you want, but are you looking to determine which sub-agents to call (in parallel) at runtime? You might want to look at writing a custom agent and implement your logic in _run_async_impl. Have a look at this example.

[boyangsvl]

We are thinking of building an agent.clone() method so that you can put multiple clones of the same agent as sub_agents. We are currently working on other items but hopefully will get there soon.

[consilience-dev]

Thanks for the reply! That's exciting, is there an issue open for this yet, or any other info? Apologies if there is, I'm still getting caught up with ADK :)

Also for this specific use case, where we may want anywhere from 1 to N sub agents to run in parallel, will agent.clone() be able to handle it?

It also begs the question, will we be able to have agents create sub_agents?