Microsoft’s New Agent Optimiser Explained (And Why It Matters to Anyone Building AI Agents)

If you have recently built an AI agent, you know the drill. Write a prompt, test it, it breaks on some edge case, tweak it, test it again, repeat. It works but it is slow and mostly guess work.
Microsoft's latest Foundry update brings something to directly address that problem: an Agent Optimiser. It automatically generates better configurations based on an evaluation of how an agent is doing, rather than a person tuning prompts and configs manually. So it’s really kind of closing the feedback loop that had to have a human sitting there and iterating by hand.”
There were a few other things that came in with the same update worth knowing if you’re working in this space: It’s a production option there, not a preview, now that it’s generally available on Microsoft Foundry. Hosted agents are expected to be generally available by July 2026, i.e. agents get a real managed runtime and don’t have to live in prototype/demo limbo. It’s more important than it sounds. A lot of agent projects die in the gap between "works in my notebook" and "runs reliably in production". A managed runtime is designed specifically to close that gap. If you are building agent-based systems today, the way forward looks pretty clear. Agent development is shifting away from manual, artisanal prompt engineering toward more of an actual ML lifecycle, evaluate,optimize, deploy, monitor, similar to how model training itself matured over the last decade.
If you want the full rundown of what shipped:
Read the full update here