Pedro Rodrigues

Pedro Rodrigues is an AI tooling engineer at Supabase and co-founder of Lisbon AI Week. He builds infrastructure that helps coding agents use databases and backend services safely, combining Supabase Agent Skills, Model Context Protocol integrations, PostgreSQL security, and systematic evaluation.

Earlier in his career, Rodrigues was a founding engineer at Ward Analytics, whose WardGraph project investigates blockchain transactions and financial risk. At SingleStore, he developed database tooling for both human developers and AI systems: he introduced its Drizzle ORM integration in December 2024 and co-authored the launch of its open-source MCP server in March 2025, giving assistants structured access to database queries and workspace operations.

At Supabase, his focus expanded from connecting agents to databases to teaching them how to operate an entire backend platform responsibly. He published PostgreSQL guidance for AI agents in January 2026, introduced Supabase’s official Claude connector in February, and launched Supabase Agent Skills in April.

The engineering ideas behind his work

• Agent experience is a product discipline. Rodrigues applies developer-experience thinking to autonomous software: agents need current documentation, explicit product conventions, reliable integrations, and workflows designed around how they actually select tools and generate code.

• MCP provides access; skills provide judgment. MCP tools let agents interact with external services, while skills supply product-specific instructions and security constraints. His preferred database workflow applies schema changes in development, runs security and performance advisors, and creates migrations after problems are resolved.

• Row-level security must extend to generated views. A PostgreSQL view can expose protected records even when its underlying tables enforce row-level security. Rodrigues uses security_invoker = true to ensure querying-user permissions govern the view, illustrating why agents need explicit safeguards alongside database access.

• Progressive disclosure has security limits. Skill metadata can defer detailed instructions until needed, but agents may skip optional reference files. Rodrigues places indispensable security rules directly inside SKILL.md and directs changing product information to maintained documentation.

• Cross-model agent evaluation must test the evaluator. He compares baseline agents, MCP-equipped agents, and agents using both MCP and skills across Claude and GPT models. His hands-on workshop also demonstrated how an incorrectly implemented grading check can misclassify correct behavior, making evaluator validation as important as the agent output itself.

Through Lisbon AI Week, Rodrigues also helps organize the Portuguese capital’s community of AI builders.