Generalize to conversations between a network of agents

[gagb]

Why is this needed:

• Current autogen can drift away from the goal ("Goodbye!"), suggest unsafe code that is unrelated to the goal. Additional, specialized agents (one for each for issue) can intervene and use conversations to fix these issues.

[sonichi]

I have an idea of implementing this via OpenAI's "functions" feature. We can define a function for each agent. For example, we can define a "goal checking" function which invokes a GoalCheckingAgent's receive method.

The benefit of this implementation is that we don't need to worry about the order of arranging the conversations etc. The agent chooses what function to call, i.e., which other agent to talk to.

Dependency: #1085