OPEN LINK ↗
// 71d agoBENCHMARK RESULT
Vercel d0 gains reliability by deleting tools
Vercel says its internal text-to-SQL agent d0 performed better after replacing a large specialized toolchain with a file-system-first setup that mainly executes shell commands in a sandbox. In Vercel’s December 22, 2025 post, the new approach improved success from 80% to 100% on five representative queries while cutting latency, steps, and token usage.
// ANALYSIS
This is a strong data point for “less harness, more model,” but the hidden variable is data quality, not just tool count.
- –The reported jump (3.5x faster, 37% fewer tokens, 42% fewer steps) suggests orchestration overhead was a bigger bottleneck than model capability.
- –Vercel explicitly says this worked because its semantic layer was already clean and well documented; messy schemas likely won’t see the same gains.
- –The benchmark covers only five queries, so teams should treat this as a design pattern to validate, not a universal rule.
- –The practical takeaway for AI teams: invest in structured files and naming discipline before adding more agent tools.
// TAGS
vercel-d0vercelagentautomationprompt-engineeringclouddevtool
DISCOVERED
71d ago
2026-03-17
PUBLISHED
71d ago
2026-03-17
RELEVANCE
9/ 10
AUTHOR
Prompt Engineering