Matt DeBergalis, CTO and co-founder of Apollo GraphQL, makes the case that AI agents should be treated as untrusted — maybe even adversarial — code running inside your firewall. In this conversation with Alex Salkever, Matt breaks down why the rush to wire agents into every enterprise system through MCP is creating a brand-new security surface, and how GraphQL's typed, governed "supergraph" model gives teams a safer way to connect agents to their APIs. It's a sharp, practical look at the collision of MCP, GraphQL, and enterprise AI from one of the people building the plumbing.

Shaun Smith is a software engineer, open source advocate, and MCP server maintainer at Hugging Face — creator of Fast Agent, the first agent framework built from the ground up around the Model Context Protocol. In this conversation from the MCP Dev Summit in New York, Shaun and Demetrios dig into why powerful models are "liquifying" open source libraries, how MCP Apps are reshaping AI interfaces, and what it means when the distribution of ideas matters more than the distribution of code.

This post compares the leading model serving runtimes for deploying production ML APIs, including TensorFlow Serving, TorchServe, BentoML, and NVIDIA Triton Inference Server. It breaks down the strengths, trade-offs, and key selection criteria—such as framework support, inference performance, infrastructure integration, and ease of use—to help teams choose the best runtime for their machine learning workloads.