Microsoft Azure outlines AI architecture choices for skills and sub‑agents

Azure chief engineer Kishorekumar Pattabiraman explains practical criteria for deciding between AI skills and sub‑agents when building conversational systems. He stresses that the primary architectural decision is whether to implement a reusable skill that runs continuously in dialogue or a sub‑agent that processes a single prompt and returns a final result. Four dimensions guide the choice: iteration model, voice fidelity, points of human intervention, and task frequency, with frequency being the clearest separator—one‑off tasks favor skills, repeatable batch work favors sub‑agents.

The articles also introduce an "AI gateway" as an evolutionary architecture pattern that isolates the rapidly changing AI layer (model updates, routing, identity, action policies, and audit) from the stable enterprise core. This gateway provides a control plane for AI components, addressing challenges such as fast‑moving model releases, evolving protocols, and new security attack surfaces like prompt injection. By routing AI traffic through this seam, enterprises can keep their main platforms stable while allowing AI capabilities to evolve independently.