Europe 2026
Combine Skills and MCP to Close the Context Gap — Pedro Rodrigues, Supabase
Overview
This talk explores how to effectively guide AI agents when they interact with complex products like Superbase. It argues that while agents are capable of many tasks, they require specific guidance to handle novel situations, security protocols, and optimized workflows. The presentation introduces the concept of "skills" as a method to provide this guidance, demonstrating how they improve agent performance and safety, particularly when integrated with tool-calling capabilities.
Who should watch
- AI Engineers building or integrating with AI agents.
- Product Managers seeking to enhance agent capabilities for their products.
- Developers looking to improve agent reliability and safety.
- Anyone interested in the practical application of agent skills and tool calling.
- Those facing challenges with agents operating on stale knowledge or missing critical product-specific information.
Key takeaways
- AI agents often struggle with tasks involving new or updated product information, security pitfalls, and require guidance on optimized workflows.
- Skills are folders containing instructions, scripts, and resources that agents can discover, with front matter for name and description, and a main skill.md file.
- Integrating skills with MCP (Multi-Call Protocol) significantly improves agent performance and task completion compared to using MCP alone.
- When building skills, avoid duplicating information; point agents to existing documentation as the single source of truth.
- Agents are often lazy to load reference files; critical, non-negotiable information should be embedded directly into the skill.md file.
- Be opinionated when designing skills; guide agents on the most effective workflows for your specific product.
- Exposing documentation via SSH allows agents to navigate it like a file system, simplifying information retrieval.
- Skills can be evaluated using frameworks like evals, similar to CI testing for code, to measure agent behavior and performance.
Notable quotes
*The bottom line is not the context, it's the guidance.*
*If you can take something from this talk when building a skill for your product is that points for your single source of truth.*
Up next · MCP & tools
Watch next
Security is not optional: insecure MCP dies in production.
Your Insecure MCP Server Won't Survive Production — Tun Shwe, LensesUnofficial community note. Prefer the recording for nuance.