If you create an agent that communicates with several users simultaneously, you need to ask yourself some questions. Your agent acts as server to many clients, the users. Each of these users will get their own account with this agent, so the agent can tell them apart. And the agent forms a user model for each of these users. It will store memories for each session with each of these users. It will solve problems for these users. The knowledge it acquires from one user can be used directly in the conversation with another user.

Can this be a single agent? A single agent that handles potentially hundreds of problems at the same time, holds the contexts of just as many conversations in its working memory, and can be at the same time outraged and deeply tranquil... err... no. It may be possible to perform such a stunt on a computer, but it would make things very much more complicated.

A single agent should communicate with a single user at a time (we presume the agent is using a simple terminal interface). If the user ends a session and returns some time later, the same agent will contact him, bringing with him all experiences of the last session. And so it will go for each user. Each user will be appointed a special agent.

But, wouldn't it be a waste of knowledge if all these agents were to stay completely unaware of all the others? Yes it would. Maybe we should let the agents talk to among themselves if they are not talking to users? Great idea! But this is future music (pardon my Dutch expression).