Nicholas Carlini, a researcher at Anthropic's Safeguards team, developed "agent teams" where multiple Claude instances work autonomously on shared codebases without human intervention. He tested this by having 16 agents collaboratively write a Rust-based C compiler capable of compiling the Linux kernel, spending $20,000 in API costs and producing 100,000 lines of code. The approach involved running Claude instances in parallel Docker containers with git synchronization, high-quality testing harnesses, and specialized agent roles. While successful at complex tasks, the method hit limitations when agents duplicated work on singular tasks like kernel compilation, requiring creative solutions like using GCC as a reference oracle.