Making Turborepo 96% faster with agents, sandboxes, and humans

This post details significant performance improvements in Turborepo's task graph computation, achieving up to 96% speedup. Key contributions include: 1. Initial performance gains from unattended AI agents, identifying optimizations in hashing (replacing twox-hash with xxhash-rust) and algorithmic improvements (replacing Floyd-Warshall with DFS). 2. Development of a new `turborepo-profile-md` crate to generate LLM-friendly Markdown profiles from Chrome Tracing JSON, enabling agents and humans to more easily parse and analyze performance data. 3. Integration of Vercel Sandboxes for testing and verification. 4. Refinements to hashing algorithms (hashing by reference instead of cloning HashMaps) and algorithmic replacements to reduce allocation pressure and improve execution speed. The post emphasizes the iterative process of combining AI assistance with traditional engineering practices, including profiling and benchmarking, to achieve these gains.