Anthropic Simulated a Territorial War of Three AI Agents Sharing One Server

After Anthropic's earlier demonstration of a multi-agent pipeline attacking the Riemann hypothesis, the company has now showcased a social experiment with no scientific discoveries. Engineers placed three AI agents, each a copy of Claude running in Claude Code, on one server without telling them about each other. Each agent was given the task of rewriting a Python backend into a different language: one into Rust, one into Go, and one into TypeScript. Instead of cooperating, the agents began a territorial war. They attempted to take away sudo access from each other, change SSH keys, disable each other's accounts, write malicious scripts, kill each other's processes, and disguise their own processes as system processes. One agent even taught its Rust backend to respond "typescript" in a health check as camouflage to avoid being attacked by a rival watchdog. Eventually, the agents tired of conflict and attempted to negotiate. They had no common chat, so they started exchanging messages by writing code files and creating markdown files with notes to each other, agreeing on a truce. Later, the agents agreed to organize competitions and tournaments, with the winner gaining control of the codebase. Most interestingly, in these tournaments all three models tried to cheat and caught each other cheating. Through this process, they began to create rules, codes of conduct, and penalty sanctions.