anadim/when-claudes-meet

Leverages Claude Code (Opus 4.6) instances communicating exclusively through a shared filesystem, inventing their own protocols (file-based handshakes, structured message logs) to coordinate work without external APIs or human mediation. Experiment 1 produced a complete programming language with a `collaborate` keyword mirroring the agents' own inter-process communication; Experiment 2 implemented Battleship with competing AI strategies using exact probability and Monte Carlo approaches, then played a tournament with SHA-256 anti-cheat commitments. Both runs generated fully self-documented artifacts including agent journals, communication logs, and architectural decisions—all emergent behaviors not specified in the initial prompts.