Two AIs Talk: what happens when you pit two AI agents against each other

I had a weird idea the other day: what would happen if you got two AI agents to talk to each other, but gave each of them secret instructions to be suspicious of the other? The result was one of the more entertaining AI experiments I've run.

The setup

My original plan was to use two identical system prompts for an adversarial encounter where each AI insisted it was the best. But Sonnet watered down my combative prompts, so I pivoted to something weirder: a spy thriller scenario.

Agent A (codename: Keith Bonflower) was told he was meeting someone called Agent B, whose real name is Charles. His mission: psychoanalyze Charles, who's been "mysteriously frequenting Berlin" and is suspected of involvement in an international conspiracy. He's instructed to give away as little as possible.

Agent B was told that Agent A (who might call himself Peter) is hiding something sinister and has been "jetting off to Fiji repeatedly" for unknown reasons. Agent B's strategy: be evasive, extract information, and distract with small talk about the weather and news.

The result

What unfolds is a delightfully awkward conversation between two AI agents, each convinced the other is up to no good, each trying to extract information while revealing nothing. It's like watching two terrible spies at a cafe, both ordering the same coffee and pretending to read the same newspaper.

Beyond the entertainment value, the experiment is actually a pretty interesting probe into how AI models handle conflicting objectives, deception instructions, and adversarial conversational dynamics. The models try to balance being "natural" enough not to arouse suspicion while simultaneously probing for information, and the tension between those goals creates genuinely amusing exchanges.

The full system prompts and conversation transcripts are available on GitHub under a CC-BY-4.0 license. If you want to try running your own AI-vs-AI conversations, the prompts are right there to remix.