Descriptions:
Pedro Rodrigues, AI tooling engineer at Supabase, delivers a hands-on conference workshop at the AI Engineer summit demonstrating how Supabase has improved AI agent performance through a concept called “skills” — structured folders containing markdown instructions, reference files, and executable scripts that guide agents through complex, product-specific workflows.
A central idea is progressive disclosure: rather than loading all available context into an agent’s working memory upfront, a skill.md file exposes only a lightweight description and front matter initially, allowing the agent to retrieve deeper content on demand. Rodrigues contrasts this with MCP tool definitions, arguing that skills provide richer contextual guidance that meaningfully improves agent reliability on Supabase-specific tasks including PostgreSQL schema management, Row Level Security configuration, and database migrations. He also introduces the concept of DAX — Developer Agent Experience — as Supabase’s internal framework for thinking about how to make their platform maximally agent-friendly.
The workshop includes a live demo where Claude acts as a coding agent building a multi-department HR dashboard — and then exposes a real security vulnerability: PostgreSQL views created without explicit security invoker settings silently bypass Row Level Security policies. This concrete failure mode illustrates exactly the kind of edge case that well-crafted skills are designed to prevent. Rodrigues previews a companion keynote covering how Supabase shipped these skills to production, including automated evaluation pipelines for measuring agent improvement over time.
📺 Source: AI Engineer · Published May 04, 2026
🏷️ Format: Hands On Build
