Data & analytics

Vercel removed 80% of its data agent's tools and made it 3.5x faster

By Vercel
Claude Opus 4.5AI SDKVercel SandboxCube

Built and shared by Vercel, not by me. Summarised from the public sources below. Check them before you rely on a number.

Who
Vercel
Category
Data & analytics
Source
Web · X
Dates
Published Dec 2025 · Found 17 Oct 2026

Vercel had spent months building an internal text-to-SQL agent, d0, with specialised tools and heavy prompt engineering. When the team replaced most of the tools with a sandbox where the agent can run bash commands over the semantic-layer files, the agent became both more accurate and faster.

What they built

d0 is Vercel's internal Slack agent for asking questions of the analytics warehouse. The new version runs Claude Opus 4.5 through the AI SDK, explores context in Vercel Sandbox, and reads the Cube semantic layer as a folder of YAML, Markdown and JSON files instead of calling many specialised tools. It is exposed in Slack through a Next.js API route.

The result

In Vercel's own test, the success rate went from 80% to 100%, average response time from 274.8 to 77.4 seconds (3.5x faster), with 37% fewer tokens and 42% fewer steps (self-reported, small test set).

Sources

Next step

Want this in
your business?

Tell me which case caught your eye. I will tell you straight whether it fits, and what it would take.

Start a conversation →