Bottom line: A team of 16 parallel Claude AI agents successfully created a complete C compiler capable of compiling the Linux kernel. The innovative "Agent Teams" system demonstrates new possibilities for autonomous language model agents, but also reveals the limits of this technology.
Researcher Nicholas Carlini from Anthropic’s Safeguards team has developed an innovative method for monitoring language models: so-called "Agent Teams". 16 parallel Claude instances created from scratch a Rust-based C compiler capable of compiling the Linux kernel.
Carlini has experimented with a new methodology in which multiple Claude instances work simultaneously on a shared codebase without requiring active human intervention. This approach significantly expands the scope of what is possible with LLM agents.
To test the system, Carlini tasked 16 agents with writing a Rust-based C compiler from scratch. This compiler needed to be capable of compiling the Linux kernel. The project required nearly 2,000 Claude coding sessions and cost $20,000 in API fees. The result is a 100,000-line compiler that successfully compiles Linux 6.9 on x86, ARM, and RISC-V architectures.
While the compiler is an interesting artifact in its own right, Carlini’s focus is on practical insights into designing control mechanisms for long-running autonomous agent teams: how to write tests that keep agents on track without human supervision, how to structure work so multiple agents can make progress in parallel, and where this approach reaches its limits.
The technical implementation works through a loop that keeps Claude on a continuous task list. Once a task is completed, the system automatically selects the next one. For parallel execution, multiple Claude instances are run in separate Docker containers with the shared repository. To prevent two agents from working on the same task simultaneously, the infrastructure uses a simple synchronization algorithm with file locks.
Source: www.anthropic.com