[Athens team] Bug: Agents frequently talk to themselves

[dimits-ts]

Even in the presence of other participants, LLM agents with the default prompt frequently refer to themselves as if they are different users.

Example from our latest experiment:

[dimits-ts]

From my own experiments, LLMs can understand that if two comments are made by the same username, then they are made by the same author. Thus, we could probably fix this by hardcoding a prompt explicitly telling LLM agents not to respond to themselves.

[dimits-ts]

Further playtesting with these kinds of prompts seems to partially succeed; Telling the agent their name on the prompt prevents them from treating themselves as distinct users. However, the agent still talks incessantly. Tested with (the free version of) Gemini.

I definitely think a small change in the prompt, which includes the agent's name, would be a good addition. The problem may be in making it dynamic when modify the agent's name.