It has become a meme recently to think about two A.I.s striking up a conversation with each other. There is indeed something spooky about this scenario — two soulless entities enacting a hollow imitation of human connection.
I first encountered this scenario when I read Isaac Asimov’s Foundation, the first parts of which were published in 1942. At one point he includes such a scenario as a kind of world-building aside.
Somewhere in the vast storage rooms of Trantor, the capital planet of the crumbling Galactic Empire, two A.I. entities are placed in storage. One is a holographic A.I. portrait of someone’s deceased wife, created many years earlier to converse with her grieving husband in imitation of her own style and personality.
The other is a similar holographic A.I. of someone’s deceased husband, created for a similar purpose. Each is programmed to respond when talked to, continuing the conversation.
In Asimov’s telling, the two framed portraits happened to be placed face to face on a shelf, where they proceed to spend several centuries, long after their creators are dead, in conversation with one another, until their futuristic batteries run down. Neither is actually alive or sentient, and the entire interaction is perfectly meaningless, yet there they are.
Whenever I think of A.I. engaging in conversation, I think back to this scene. There is something incredibly sad about it, and it has continued to haunt me to this day.
The Asimov reference is a good catch. I went back to those Trantor passages after reading this, and what struck me is how little the two entities actually want from each other. They end up stored together because of an administrative decision, and neither seems curious about the other.
Where I have seen this play out for real is in tool calling. Two agents handed the same task will pass it back and forth, each assuming the other holds the context. I lost an afternoon to a loop like that in March, two models politely rephrasing the same question until I capped the turns.
The fix for us was giving one side a hard budget and a stop condition. That is also why I prefer APIs with visible per-call cost when wiring agents together. pixapi.ai is what I use for the image side of those pipelines.