Microsoft has made Routines generally available in Foundry Agent Service, giving developers a managed way to run AI agents on schedules, at specific times, or in response to external events without building custom automation infrastructure. The feature moves agents beyond reactive chat interactions into proactive, unattended workflows that can monitor systems, respond to business events, and continue tasks over time.
Announced by product manager Linda Li, Routines handles the underlying plumbing — schedulers, event listeners, webhooks, queues, authentication, run history, and monitoring — so teams can focus on agent capabilities rather than infrastructure. The service supports three trigger types: timer (once at a future date), recurring (on a cron schedule), and event-based (when something changes in a connected system). Initial event integrations can react to GitHub issues and messages posted in Microsoft Teams channels.
The trigger types map to distinct operational patterns. Timer routines suit one-off moments, such as getting ready for a release date, nudging people ahead of deadlines, or queuing up later check-ins. Recurring routines fit repeating work like daily digest generation, weekly reporting cycles, compliance audits, backlog triage, and ongoing operational monitoring, so teams no longer need to run a separate scheduler and invocation service next to the agent itself. Event-based routines let agents react when something changes in a connected system rather than when someone remembers to ask: Foundry receives the event through the configured connection, forwards its payload to the agent, and the agent can reason over it, use its tools, and take the next action — for example, a GitHub agent triaging an issue as soon as it arrives, or a Teams agent detecting a support request, retrieving context, and initiating a response without a separate chat interaction.
Routines also cover work that spans more than one execution. An agent could kick off a data export, pause for an approval step, trigger a deployment, or hit an API that takes minutes to return. The reminder tool (preview) lets a Hosted Agent schedule itself to run again after a delay on the same conversation, preserving context: the agent starts a long-running task, sets a reminder, gets re-invoked when it fires, checks status, and repeats until the work is done — no polling service or separately persisted conversational state required. On the identity side, developers choose per routine between creator identity, which acts with the creator's delegated access within that person's permission boundaries, and agent identity, which runs as an independent agent with its own configured permissions in Microsoft Entra ID. Foundry manages the underlying authentication and token exchange, so agent code stays free of identity wiring while automated activity remains governable and auditable.
Confirmed
- Routines in Foundry Agent Service is generally available as of the announcement date.
- Three trigger types: timer, recurring (cron expressions with time zone support), and event-based.
- Initial event sources: GitHub issue events and Microsoft Teams channel messages.
- Reminder tool (preview) lets a Hosted Agent schedule itself to run again after a delay, preserving conversation context.
- Two identity models: creator identity (delegated user access via OAuth) and agent identity (independent Entra ID permissions with key-based or managed identity auth).
- SDK support: Python, JavaScript, .NET, REST API, and Azure Developer CLI.
- Management via Microsoft Foundry portal.
Unknown
- Pricing model for Routines executions (separate from agent compute costs).
- SLA or throughput limits for event-based trigger delivery latency.
- Full list of planned event-source integrations beyond GitHub and Teams.
- Whether reminder tool preview has a timeline for general availability.
- Maximum routine duration or retry policies for long-running agent tasks.
Our take
Routines target a real operational gap: before an agent can act on its own, someone has to build the scheduler, event handling, authentication and run history around it. Moving that layer into Foundry turns agent invocation into a platform feature rather than a custom project. The choice between creator and agent identity suits enterprise governance, while the preview status of the reminder tool shows that multi-step autonomous loops are still maturing. Pricing, which the announcement does not detail, will decide how widely teams adopt it.
