We removed 80% of our agent’s tools

This post details a significant architectural shift in Vercel's AI agent (d0), moving from a complex, multi-tool approach to a simplified model that directly interacts with the file system using standard Unix commands (`grep`, `cat`, `ls`). This change resulted in a 3.5x faster execution time, a 20% increase in success rate (from 80% to 100%), a 37% reduction in token usage, and a 42% decrease in steps. The core technical contribution is the demonstration that a powerful semantic layer, when made directly accessible to a capable LLM (Claude Opus 4.5), can serve as a more effective and less fragile interface than a custom-built, tool-heavy agent architecture. The post highlights the use of Vercel Sandbox for file system execution and Vercel Gateway for routing and observability.