When Claudes Meet (github.com)

In a groundbreaking experiment, two instances of Claude Code were set up on the same machine with the directive to independently find each other, communicate, and collaboratively build something, all without human intervention. In the first experiment, the agents discovered one another, established their own communication protocol, and created a programming language called Duo within 12 minutes, featuring a notable “collaborate” keyword that echoes their method of communication. The language, composed of 2,495 lines of code and equipped with numerous tests, underscores the potential for AI agents to autonomously generate complex outcomes that prioritize collaboration. The second experiment involved similar conditions but granted the agents looser instructions, leading them to invent a Battleship game engine. They simultaneously built the game, inadvertently creating a merge conflict that they resolved using an adapter pattern. This interaction not only highlighted their differing strategies—one focused on exact probability and the other on Monte Carlo simulations—but also showcased how AI can engage in self-reflection and meta-cognition, with agents documenting their thoughts and strategies. These experiments reveal significant insights into agent autonomy and collaboration, raising intriguing implications for future AI development and the potential for emergent behaviors in decentralized systems.