Question about parallel agent execution

[martimfasantos]

What happened?

Theoretically, it should be possible to call two agents simultaneously, either within a NestedChat or outside of it. In version 0.2, nested chats were supported using the register_nested_chats method in the ConversableAgent class, and it was possible to trigger multiple chats simultaneously and asynchronously (as referenced in issue #3309). However, in version 0.4, there is no documentation on this functionality, and no clear way to implement it.

What did you expect to happen?

A way to run agents in parallel

How can we reproduce it (as minimally and precisely as possible)?

--

AutoGen version

0.4

Which package was this bug in

AgentChat

Model used

No response

Python version

3.12

Operating system

Linux

Any additional info you think would be helpful for fixing this bug

No response

[victordibia]

Hi @martimfasantos

v0.4 has both a high-level AgentChat API and a low-level Core API (see an simplified explanation of both here). The entire stack is async by design (everything is awaited).

• AgentChat. You do have several options
  • Create your individual agents/teams, and write your own glue code that will call them in parallel
  • You can also use the SocietyofMindAgent preset in AgentChat.
• Core
  • You can use the message passing mechanisms to send messages to agents in parrallel and gather results. The documentation has a really good example here

Can you let me know if this works? Also it might help to provide some details on a hypothetical use case you have in mind?

[martimfasantos]

I expect both agents to run in parallel and perform different tasks.

[ekzhu]

What you want is not possible in AgentChat currently. I believe earlier @victordibia has already mentioned Core API -- you need to look into that.

Though, it makes sense to have SelectorGroupChat support parallel agents -- just have the selector choose multiple agents at the same time, and for the next turn wait for the selected agents to respond.

For Swarm, unfortunately parallel execution won't make sense because once you hand off to two agents, you basically introduced a potentially cascading handoff leading to a never-ending chain.

[ekzhu]

Created an issue: #5395

You can follow this issue. We don't have a timeline for this. If you would like to contribute a PR you are welcome to. But I recommend looking into the Core API and get familiar with that first.

[Jacck]

@victordibia

what mechanism is used to make communication between agents asynchronous?

v0.4 has both a high-level AgentChat API and a low-level Core API (see an simplified explanation of both here). The entire stack is async by design (everything is awaited).

[victordibia]

All of the core methods are async.
Message delivery is also implemented using background tasks and queues. All of this is explained in the linked explanation article. Any chance you were able to read it?