Will running multiple instances of an agent in parallel cause "interference"?

[alex-treebeard]

Hi there,

I'm integrating Autogen with Inspect which will let us run many instances of an agent in parallel for running evals.

Specifically I'm creating our own subclass of MagneticOneGroupChat

One of my concerns is a lot of the message passing seems to refer to agents as "topic", making me wonder if we will get subtle logic errors where different instances of our agent are handling each others' messages.

My q's:

1. is this a valid concern?
2. If so, how could we mitigate this? (e.g. giving each agent instance a unique name that is invisible to the LLM)
3. Are there other issues relating to large-scale evals I should be aware of?

[ekzhu]

By default, each team uses its own agent runtime and message queue so different teams don't interfere with each other.

Within the same team, each agent has its own unique topic type. So message published to each agents own topic type won't interfere with other agents. Messages published to the group topic type will be observed by all agents.

My guess is if you are using default settings, nothing to worry about if you are calling the team from outside.