Making Turborepo 96% faster with agents, sandboxes, and humans
91% faster. That's what Turborepo 2.9 achieved in 8 days by mixing AI agents, clean benchmarking, and old-school engineering.
Why it matters
Vercel shipped a major Turborepo performance release powered by an AI-assisted optimization workflow. The story reveals both the real capabilities and hard limitations of coding agents—and how human judgment remains essential for shipping production-grade software at scale.
The key facts
10 to knowTurborepo 2.9: 91% faster on 1,000+ package monorepos (8.1s → 0.716s)
Time to First Task: 11x faster on largest repos
81-91% faster depending on repo size and complexity
8 AI agents spawned for unattended optimization baseline; only 3 of 8 produced shippable wins
20+ performance PRs generated in 4 days using human-agent iterative loop
Three optimization categories: Parallelization (largest wins), Allocation Elimination, Syscall Reduction
Vercel Sandboxes used for noise-free benchmarking (vs. noisy laptop results)
Key technical wins: Stack-allocated git OIDs (15% self-time reduction), Syscall elimination (35% reduction on cache fetch), HashMap move vs. clone optimization
Agent limitations documented: no regression tests written, hyperfixation on first ideas, meaningless microbenchmarks, failed to dogfood Turborepo itself
Markdown profile format dramatically improved agent output quality vs. JSON trace format
Go to the source
Vercel Blogvercel.com
Publisher excerpt: is now 81-91% faster to compute its task graph in our repositories, scaling with repo size. On our 1,000+ package monorepo, now feels instant. Time to First Task is now 11x faster.Turborepoturbo run After testing my changes with some open source Turborepos and asking Vercel customers to try canary…