Passing on control to your AI coding agent team entirely?

Passing on control to your AI coding agent team entirely?

Anthropic researcher Nicholas Carlini conducted a stress test of their Claude Opus 4.6 model by deploying 16 parallel AI agents to build a C compiler in Rust from scratch, aimed at compiling the Linux kernel (https://www.anthropic.com/engineering/building-c-compiler). Over approximately two weeks and nearly 2,000 Claude Code sessions, the agents — running without active human intervention, inside a harness, test suites, and CI that Carlini designed and kept refining — produced a 100,000-line compiler that builds the Linux 6.9 kernel on x86, ARM, and RISC-V. The experiment cost around $20,000 in API fees. Carlini’s write-up is refreshingly candid about the limits: the compiler still calls out to GCC for the 16-bit x86 boot stage and for assembling and linking, doesn’t build every project, and even fully optimized emits slower code than GCC with optimizations disabled. One expensive, deliberately extreme data point — but a striking one for what coordinated agent teams can produce.

Adopting this approach would be a nice extension for my pet project where I’m getting good results with a team of Kiro Subagents. I recently wrote about this in https://linkedin.com/posts/stefanchristoph_kiro-activity-7426663959514161152--4-g Having a self-managed orchestrator who runs the team while I’m focusing on other things sounds interesting. At least as an experiment.

💬 Question to you: Have you experimented with parallel AI agent architectures in your own projects? Obviously at smaller scales - quite some budget being burned by the original experiment.

And if you’re using Kiro: have you tried running similar subagent patterns with Claude Code instead? I’d love to hear about your experiences and what worked (or didn’t) for you.

#AI #AgenticAI #SoftwareEngineering #Claude #Kiro

• AI
• AIAgents
• SoftwareEngineering