Nicholas Carlini, a researcher at Anthropic, shares his experience with a new approach to managing language models called "agent teams." He deployed 16 instances of their model, Claude, to collaboratively write a Rust-based C compiler capable of compiling the Linux kernel. Over nearly 2,000 sessions and $20,000 in API costs, the agents produced a 100,000-line compiler that can build Linux 6.9 across multiple architectures, including x86, ARM, and RISC-V. This experiment highlights both the potential and challenges of employing autonomous agent teams for complex tasks. Carlini's method involved creating a continuous loop for each Claude instance, allowing them to work independently without human intervention. He implemented a synchronization mechanism to prevent agents from overlapping on tasks. Each agent could claim a task by writing to a shared directory, and they would pull and merge changes from one another as they progressed. Despite the success, Carlini acknowledges that this approach is still in its early stages. He hasn’t yet established communication protocols between agents or an orchestration agent, leaving each Claude to decide its next action based on the most obvious problem at hand. Key lessons emerged from the project. Writing high-quality tests was critical, as any flaws in the verification process could lead Claude to solve the wrong problems. The author learned to anticipate the limitations of the language model, such as context window pollution and time blindness, adjusting the environment to accommodate these constraints. For instance, he minimized clutter in output logs and implemented a sampling strategy to keep progress manageable. When compiling the Linux kernel, however, the agents faced challenges due to the task's complexity, often hitting the same bugs. To address this, they used GCC as a reference to ensure consistency and guide their progress.