We Removed 80% of Our Agent's Tools
Stripping a text-to-SQL agent down to a single bash tool produced a 3.5x speedup, 100% success rate, and 37% fewer tokens—proving that simpler agent architectures outperform elaborate tooling.
The Problem
Vercel built a sophisticated text-to-SQL agent with specialized tools: schema lookups, query validators, error recovery mechanisms, and dimensional attribute retrievers. Each edge case demanded another tool. Maintenance grew unbearable as models updated.
The Solution
They stripped the system to a single capability: bash execution. The agent now uses grep, cat, find, and ls to navigate a well-documented semantic layer (YAML and JSON files containing dimension definitions and join relationships).
Results
| Metric | Before | After | Improvement |
|---|---|---|---|
| Execution time | 274.8s | 77.4s | 3.5x faster |
| Success rate | 80% | 100% | +20pp |
| Token usage | ~102k | ~61k | 37% fewer |
| Processing steps | ~12 | ~7 | 42% fewer |
Technical Stack
- Model: Claude Opus 4.5 via AI SDK
- Execution: Vercel Sandbox
- Infrastructure: Next.js API route with Slack integration
- Data: Cube semantic layer as structured files
Key Lessons
Trust the model's reasoning. Advanced models perform better when decision-making isn't artificially constrained by custom retrieval logic.
Quality foundations matter. Success depends on well-structured, documented data—not clever tooling. The semantic layer already contained everything the agent needed.
Build for emerging models. Design architectures that let you delete yesterday's assumptions. The winning strategy is subtraction, not addition.
"Don't fight gravity." Unix tools remain powerful abstractions.
Connections
- context-engineering-for-ai-agents-with-langchain-and-manus - Manus discovered the same counterintuitive insight: "As models get stronger, we should be getting out of the model's way." Both teams found that removing scaffolding beats adding it.
- building-effective-agents - Anthropic's "simplicity first" principle in action. Vercel's stripped-down agent exemplifies resisting complexity and letting the foundation (well-documented files) do the work.
- the-importance-of-agent-harness-in-2026 - Validates the "design for deletion" strategy. Manus rewrote their harness five times; Vercel deleted 80% of their tools. Lightweight beats elaborate.