Microsoft launches hosted agents in Foundry with secure sandboxes

Microsoft Foundry Agent Service is a fully managed platform for building, deploying, and scaling AI agents. Since it is a fully managed service, Microsoft handles the hosting, scaling, identity, observability, and security of the agents. Today, Microsoft announced the public preview of hosted agents within the Foundry Agent Service.

This new hosted agents offering will help enterprises run production-grade AI agents with dedicated per-session isolation, persistent file system states, built-in identity support, and more. This service is a complete refresh of the version Microsoft previewed during the Ignite conference last year.

With hosted agents, every agent session has its own dedicated VM-isolated sandbox with a persistent file system. Microsoft highlighted that this is not just process isolation or a limited code-execution container, but production-grade hypervisor isolation at cloud scale.

Microsoft highlighted the following core capabilities of hosted agents in this preview:

• Fast Starts and Efficiency: Hosted agents support predictable cold starts, allowing sessions to spin up quickly in custom environments. They can also scale to zero when idle, ensuring customers do not pay for inactive agent time.
• Persistent Continuity: When a session resumes, its files, disk state, and session identity remain intact, allowing the agent to continue exactly where it left off.
• Advanced Security: The service supports custom isolation keys and BYO VNet support for routing outbound traffic through a customer’s virtual network.
• Controlled Deployments: Built-in versioned agent endpoints allow for controlled deployments and rollouts.

When developing hosted agents, developers can use their preferred harness, framework, and environment, including LangGraph, Microsoft Agent Framework, Claude Agent SDK, OpenAI Agents SDK, and GitHub Copilot SDK. Developers can also define custom environments through Dockerfiles and deploy using a single azd deploy command. As expected, these hosted agents can be powered by models from OpenAI, Anthropic, Meta, Mistral, and others.

Excited to announce the new preview for Microsoft Foundry Agents 🎉! You can now build, run, and deploy your agent using any model, any framework, any harness in the cloud 🧑‍💻 - check out the demo below

This is not just any cloud compute environment; it's an agent-optimized… pic.twitter.com/mTIK6EnAlA

— Jeff Hollan (@jeffhollan) April 22, 2026

Along with hosted agents, Microsoft also announced the public preview of Toolbox, a unified way to configure and manage tools across frameworks with built-in authentication, OAuth passthrough, and observability for each tool call. Additionally, the new Memory service brings managed long-term memory directly into the Foundry Agent Service, enabling agents to retain context across sessions without requiring external databases.