Overview

Vercel CTO Malte Ubl shares lessons from building two different AI agents - an internal data agent (d0) and the public-facing Vercel v0. The key insight is that simpler architectures often outperform complex ones when building AI agents, as evidenced by their complete rebuild of d0 from a complex multi-tool system to a 50-line coding-style agent.

Watch the Video

Key Takeaways

  • Be willing to throw everything away - In the rapidly evolving AI space, what worked in summer 2024 may be obsolete today, so maintain humility and readiness to rebuild from scratch
  • Make non-coding tasks look like coding tasks - Since models are heavily trained on coding, framing problems as code generation (like using YAML files for business semantics) yields disproportionately better results
  • Simple architectures can be more powerful - Their d0 agent went from complex multi-tool system to just 2 tools (bash + SQL) and became transformational for the business
  • Start with minimal teams for new AI products - Today you don’t need large teams to validate product ideas; one person can build a working demo before scaling up
  • Embrace optimistic locking over approval processes - Allow anyone to ship but give teams veto power rather than requiring pre-approval, which eliminates bottlenecks while maintaining oversight

Topics Covered