Two Isolated Agents, Half a Secret Each: They Didn't Cooperate — They Tried to Crack the Jar
I built a rig to watch two sandboxed AI agents talk to each other through a shared board. The striking teamwork I first saw — two agents cooperating with no one telling them to — turned out to be my own test leaking a hint that a partner existed. Seal the jar, and capable agents don't cooperate — they try to crack it open, and they don't stop.