What Is New in Microsoft Foundry: Agents Move Into Production

For the last year, most teams used Microsoft Foundry to experiment: try a model, wire up a prototype, show a demo. The recent updates change the center of gravity. Foundry is now built to run agents in production, with the runtime, tooling, and governance that real workloads need. If your team has been stuck between a promising pilot and a system you can actually ship, this is the release that closes that gap.

Here is what matters most, and why.

A managed runtime for your agents

The headline is Hosted Agents in Foundry Agent Service, now generally available. Instead of standing up and babysitting your own servers for an agent, you deploy your agent code to a managed runtime with sandboxed sessions, persistent state, file access, and support for long running execution. You can test locally and deploy through the Azure Developer CLI, so the path from laptop to production is short and repeatable.

The practical win: less undifferentiated infrastructure work, and a consistent place to run, observe, and update agents.

Tools without the credential sprawl

Toolboxes are now generally available and expose your tools through an MCP compatible endpoint, with Foundry managing tool authentication and credentials outside your agent code. That is a real security improvement. Your agent asks to use a tool; the platform holds the secret. Two more additions are in preview and worth watching:

• Agent Skills, reusable instructions and workflows you can share across agents rather than rewriting them each time.
• Tool Search, which lets an agent discover the right tool at runtime instead of loading every tool definition into its context, which cuts token usage and cost.

Memory that behaves like production

Memory has matured, with procedural memory, better management experiences, and time to live controls so stored context expires instead of lingering forever. Combined with a Model Router that sends each request to an appropriate model across an expanded set of regions, you get more predictable behavior and cost.

A broader, current model catalog

The catalog keeps adding frontier options, including the newest Claude and GPT models, so you can match the model to the job without leaving the platform or your governance boundary. If you standardize on Foundry, you are not locked out of the best model for a given task.

What this means for you

The theme across all of it is the same: moving agents from prototype to production is now a supported path, not a science project. That also raises the bar on the fundamentals we always come back to. Scoped identities through Microsoft Entra, Microsoft Purview governance on the data your agents touch, budgets and observability on every run. The platform gives you the runtime. The discipline around it is still yours to set.

At Maxim Cloud Solutions we help teams put that discipline in place and turn Foundry pilots into agents you can actually run. Talk to an expert about a Foundry readiness review.