Foundry hosted agents enter public preview; Microsoft publishes MSFTMechanics demo videos
- Microsoft on April 22 put hosted agents in Foundry Agent Service into public preview and paired the launch with a new Microsoft Mechanics video showing how developers deploy agents from local code. - The refreshed preview swaps in per-session sandboxed compute, dedicated Microsoft Entra identities, stable agent endpoints, and scale-to-zero billing; Microsoft says older preview deployments must be redeployed before May 22, 2026. - The release extends Microsoft’s push to make Foundry its managed platform for enterprise AI agents, after Agent Service reached general availability earlier in 2025. (learn.microsoft.com)
Microsoft on April 22 moved hosted agents in Foundry Agent Service into public preview and published a Microsoft Mechanics demo showing how to deploy them from local development into production. (devblogs.microsoft.com) (youtube.com) Hosted agents are Microsoft’s managed way to run AI agent code after developers package it as a container image and push it to Azure Container Registry. Foundry then provisions compute, assigns a dedicated endpoint, and gives the agent its own Microsoft Entra identity. (learn.microsoft.com) Microsoft said the new preview is a refresh of an earlier hosted-agent preview first shown at Ignite, with a different backend and a session-based model. Each session gets its own isolated sandbox and persistent file system instead of sharing a project-level runtime. (devblogs.microsoft.com) (learn.microsoft.com) In plain terms, the service is aimed at the messy parts of running agents in production: starting containers, handling identity, saving state, scaling capacity, and tracing behavior. Microsoft’s documentation says developers can still use their own code and frameworks, including Agent Framework, LangGraph, Semantic Kernel, or custom stacks. (learn.microsoft.com) The Microsoft Mechanics video centers on that production handoff. Its demo walks through local deployment, scoped access controls, publishing into Microsoft 365, code-level tracing, and performance management for a sales-meeting-preparation agent. (youtube.com) Microsoft’s launch post highlighted three infrastructure claims: per-session hypervisor-isolated sandboxes, filesystem persistence across idle periods, and scale-to-zero economics when an agent is inactive. The migration guide says compute now deprovisions after 15 minutes of inactivity and resumes with persistent storage intact. (devblogs.microsoft.com) (learn.microsoft.com) The preview also changes how agents are addressed and secured. Microsoft says each agent now gets its own endpoint URL and dedicated Entra identity at deploy time, replacing the older model that routed requests through a shared project endpoint and shared project managed identity. (learn.microsoft.com) For developers, Microsoft has already published a quickstart that uses the Azure Developer CLI or a Visual Studio Code extension to initialize, test, and deploy a hosted agent. The sample uses web search and optional Model Context Protocol tools, and the docs require Azure Developer CLI 1.24.0 or later. (learn.microsoft.com) The migration deadline is concrete. Microsoft says deployments built on the initial public preview backend will not be migrated automatically and will be supported only until May 22, 2026. (learn.microsoft.com) That puts the April 22 launch in two buckets at once: a new public preview for teams starting fresh, and a forced rebuild for teams that tested the older hosted-agent stack. Microsoft’s message in both the docs and the demo is that Foundry is now where those agents are supposed to live. (devblogs.microsoft.com) (youtube.com)