Most agents I see in demos sit there until a human types something. That's a chatbot with extra steps.
Microsoft just made Routines generally available in Foundry Agent Service. It's a managed answer to "how do I make my agent run at 6 AM without me?"
A routine can fire three ways. A timer runs the agent once at a set date and time. A recurring routine uses a cron expression. And an event trigger starts the agent when something happens, with GitHub issues and Teams messages supported today.
Before this, you built the plumbing yourself. A scheduler, a queue, some identity glue, and a run history table you'd forget to clean up. Foundry now handles all of it, and it works through the Python, JavaScript and .NET SDKs or a plain REST API.
The identity choice is the part I'd read twice. Each routine runs either as the creator's identity (with their delegated permissions) or as the agent's own Microsoft Entra identity. If the person who built it leaves the company, a routine tied to their account is a problem waiting for a calendar reminder. An agent identity gives you something you can scope, audit and keep alive without anyone's login.
SpaceX Got Wall Street’s Attention. These 7 Stocks Could Follow.
SpaceX grabbed Wall Street’s attention. Now investors are looking for what could come next.
This free report names seven space stocks positioned across the industry’s fastest-growing markets and explains how each company makes money, what could move its shares and which risks could bring it back to Earth.
There's also a new reminder tool, still in preview. A hosted agent can schedule its own follow-up run after a delay and keep the conversation context. Think of an agent that kicks off a deployment, then checks back in 20 minutes instead of sitting there polling.
That pairs with something Microsoft shipped in preview around the same time. Long-running resilience lets a hosted agent's work continue after a request disconnects and recover if the hosting process dies. It uses persisted inputs, lease-based recovery and stream replay. Scheduled agents that run for a while will crash sometimes, so this matters more than it sounds.
Two easy ones to start with. A Monday morning agent that summarizes the weekend's open GitHub issues before standup. A nightly agent that checks a tagged set of resources and files a ticket when something drifts. Both used to mean a Function app and a timer trigger someone had to babysit.
One catch. The announcement doesn't give pricing, regional limits or usage caps, so I'd check the docs before you promise anyone a production rollout. And the event triggers are just GitHub and Teams for now, which is a short list.
Still, this is the boring infrastructure that turns an agent from a demo into a coworker. AWS covered a lot of the runtime side lately with AgentCore (I wrote about the cold start rebuild last Wednesday). This is Azure's turn on the scheduling side.
If you're working toward an Azure AI cert, learn the difference between creator identity and agent identity now. Who an automated process runs as, and what permissions it holds, is the kind of question Entra-heavy exams like to ask.
And if you've got an agent you currently trigger by hand every morning, that's your first routine.
Keep Learning and Building.