Overview
Microsoft Foundry Hosted Agents is reaching General Availability in early July 2026. This is the managed runtime layer for enterprise AI agents on Azure — giving organisations a production-grade hosting environment with sandboxed execution, persistent state, filesystem access, and direct deployment to Microsoft Teams and Microsoft 365 Copilot.
For organisations running on Microsoft 365, this is the deployment path that puts your agents in front of every user in your organisation with zero additional distribution infrastructure.
The 3-Layer Architecture
Layer 1: Your Agent Code
You write the agent logic using the Azure AI Agent SDK — defining tools, memory, instructions, and orchestration. Your code does not change based on how it is hosted. The same agent logic that runs locally deploys to Foundry without modification.
Layer 2: Foundry Runtime
The Foundry runtime manages execution — sandboxed sessions, state persistence, tool execution, memory management, and observability. You do not manage servers, containers, or scaling. The runtime handles all infrastructure concerns and provides the isolation guarantees that enterprise security teams require.
Layer 3: Publish Surfaces
With one deployment command, your agent can be published to Microsoft Teams as a bot and to Microsoft 365 Copilot as a plugin. Every user in your Microsoft 365 tenant gets access without any additional installation or configuration from IT.
Why Sandboxed Sessions Are the Real Differentiator
Sandboxed sessions give each user interaction an isolated execution environment. The agent can read and write files, execute code, and call APIs — all within a sandbox that is torn down when the session ends. This makes it safe to give agents filesystem and code execution access in enterprise environments where data isolation is non-negotiable.
Compare this to running agents on a shared serverless function where state can bleed between invocations, or on a dedicated container where you manage the isolation yourself. Foundry handles the sandbox lifecycle automatically.
5 Concrete Use Cases
- Code agent: Reviews pull requests, writes tests, answers codebase questions — deployed to Teams so developers can mention it in engineering channels
- Data agent: Queries databases, generates reports, explains anomalies — accessible via Copilot for analysts without SQL access
- Document agent: Processes contracts, extracts clauses, compares versions — integrated into document approval workflows
- Ops agent: Investigates alerts, runs diagnostics, escalates incidents — deployed to the on-call Teams channel
- Workflow agent: Orchestrates multi-step business processes, handles exceptions, and maintains audit trails
vs Self-Hosting on AKS, Lambda, or Cloud Run
Self-hosting gives you more control over the runtime environment but requires you to build and maintain container images, manage scaling, handle session isolation yourself, and build the Teams and Copilot integration from scratch. Foundry Hosted Agents removes all of that operational overhead in exchange for running within Foundry's constraints and Azure's geographic footprint.
Key Takeaways
- Foundry Hosted Agents GA is the production runtime for enterprise agents on Azure — available early July 2026
- Sandboxed sessions enable safe filesystem and code execution access without custom isolation infrastructure
- One-click deployment to Teams and M365 Copilot covers the full Microsoft 365 user base
- For Microsoft 365 organisations, this is the lowest-friction path to enterprise-wide agent deployment
- The trade-off is Foundry lock-in — porting agents to another runtime requires rebuilding the hosting layer
