Microsoft Foundry hosted agents: lessons learned building a production multi-agent service

This article details the author's experience building a production multi-agent service using Microsoft Foundry hosted agents, which went GA in July 2026. It covers hosting decisions, protocol choices (Responses vs. Invocations), and key lessons for bursty, scheduled, or long-running agent workloads. The author explains how hosted agents provide VM-isolated sandboxes per session with a 15-minute idle timeout and 30-day max session lifetime, preserving session contents across idle gaps. The identity story is highlighted, with each agent getting a dedicated Microsoft Entra identity for security. The article also contrasts hosted agents with always-on container replicas, emphasizing cost savings and compute efficiency. Practical advice includes sizing compute per agent, avoiding in-process callbacks for clean migration, and using Invocations over Responses for task-oriented agents.