The team at Vercel faced significant challenges with their initial AI agent, d0, designed for translating natural language into SQL queries. After investing months in building a complex system with multiple specialized tools and heavy prompt engineering, they found it to be slow, fragile, and requiring constant maintenance. The agent achieved an 80% success rate but ultimately fell short in reliability, prompting the team to rethink their approach. By stripping down the agent to just one tool—executing bash commands—they created a file system agent. This new architecture allowed the AI to directly access files and process data using standard Unix commands like grep and cat. As a result, the agent not only achieved a 100% success rate but also completed tasks significantly faster—141 seconds compared to 724 seconds with the old system—and used 37% fewer tokens. The team realized they were over-engineering solutions to problems the model could handle on its own and learned that less can be more when it comes to AI tools. Key lessons emerged from this experience. The existing semantic layer provided clear, structured documentation that made it easier for the model to reason and generate queries without additional constraints. The team emphasized the importance of starting with a simple architecture and ensuring good data context rather than overcomplicating with unnecessary tools. They also noted that models like Claude Opus 4.5 are evolving rapidly, and developers should anticipate future capabilities rather than just focusing on current limitations.