Microsoft's Foundry Agent Service just launched a public preview of its hosted agents platform, giving developers a dedicated compute environment for building and running AI agents. The killer feature? Each agent session gets its own VM-level sandbox with a persistent file system, and when you scale down to zero instances, your disk state is preserved—and you don't get billed for the downtime.
This isn't a walled garden. The platform is framework-agnostic, so you can build with LangGraph, Microsoft's own Agent Framework, Claude Agent SDK, OpenAI Agents SDK, GitHub Copilot SDK, or pretty much anything else. Define your environment with a Dockerfile, deploy with a single azd deploy command, and pick your model from a list that includes OpenAI, Anthropic, Meta, Mistral, and others.
Alongside the core platform, Microsoft rolled out a suite of companion services. Toolbox provides a unified way to manage tool calls and supports the MCP (Model Context Protocol) for integration. Memory offers cross-session long-term storage without needing an external database. And Microsoft Agent Framework v1.0 is now official, merging Semantic Kernel and AutoGen into a single toolkit. For observability, the whole stack is instrumented with OpenTelemetry for end-to-end tracing, and there's even an AI Red Teaming Agent built in for adversarial testing.
On the enterprise side, each agent gets a unique Entra Agent ID, supports on-behalf-of (OBO) user operations, and enables continuous auditing. Network integration includes VNet support, while security policies cover data loss prevention (DLP) and Responsible AI guardrails. And when you're ready to ship, you can publish your agent directly to Microsoft Teams and M365 Copilot with one click.