Descriptions:
The CTO of Vercel presents hands-on lessons from building two production AI systems: DZero (also written “d0”), an internal text-to-SQL agent that answers business questions in Slack using Snowflake data, and V0, Vercel’s public-facing AI app generation tool. The talk is notable for its candor about a complete architectural rebuild — and the reasoning that forced it.
The original DZero agent used a conventional tools-in-a-loop architecture with many specialized tools and extensive prompting. It worked, but felt uninspired. After discarding it entirely, the team rebuilt using an approach inspired by coding agents: a YAML file describing the business semantics of every Snowflake column in plain English, two tools (bash and SQL execution), and roughly 50 lines of total agent code. A companion Vercel blog post titled “All You Need Is the File System and Bash” documents the philosophy. The result is described as transformational for business users, enabling queries like identifying which S&P 500 CTOs have active private Vercel accounts.
A broader theme is the pace of obsolescence in agent design: what was best practice in summer 2025 may already be the wrong approach in early 2026, as more capable models allow simpler, more declarative architectures to outperform complex hand-crafted systems. The talk also covers Vercel’s competitive positioning — arguing that reliable production infrastructure remains a durable need regardless of how applications are generated — and addresses the question of observability and “self-driving infrastructure” as agents take on more DevOps responsibilities.
📺 Source: The Pragmatic Engineer · Published March 06, 2026
🏷️ Format: Workflow Case Study
