We removed 80% of our agent’s tools

The article describes the evolution of an internal text-to-SQL agent named d0, initially built with complex tools and extensive prompt engineering, which was found to be fragile and resource-intensive. By simplifying the agent to execute arbitrary bash commands and allowing direct access to files using Unix tools like grep, cat, and ls, the team achieved a 100% success rate, reduced steps and token usage, and faster response times. This shift highlights the effectiveness of leveraging existing file systems and trusting the model's reasoning over building elaborate custom tools. The new architecture relies on the AI model Claude Opus 4.5, utilizing Vercel Sandbox for context exploration and a Cube semantic layer for documentation, emphasizing the importance of good data documentation and simplicity in agent design. The experience underscores the value of starting with a simple architecture and investing in well-structured data, suggesting that the rapid improvement of AI models may outpace the need for complex tooling.