Summary

This technical presentation focuses on designing secure and robust MCP (Multi-Capability Provider) servers for agentic AI systems, emphasizing the unique challenges in creating interfaces optimized for AI agents. The speakers, from Lenses, highlight three key dimensions of difference between human and agent interactions: discovery, iteration, and retry mechanisms, with a particular emphasis on potential security vulnerabilities inherent in these processes. The practical takeaway is that poorly designed MCP servers increase security risks, and developers must carefully consider how agents interact with interfaces, paying special attention to potential attack surfaces like tool descriptions and conversation histories.