Exquisite Evals for Skills & MCPs

I built the javadocs.dev MCP server to help Java, Kotlin, and Scala developers give their agents more efficient access to documentation for the libraries they use. But until recently I didn't know if it was better than what base models, or base models with web search/fetch, use. I created a custom eval framework to provide verifiable evidence that the javadocs.dev MCP is up to 5x more token efficient than what the standard agents provide and provides correct answers, unlike the base models. This session will share the learned best practices for building and implementing evals that can provide the feedback you need to improve your agent skills & MCPs.