What Microsoft announced at Ignite
On 17 November Microsoft published Azure update AZ580753. The update puts Azure SRE Agent into Microsoft Teams as a public preview. Engineers can start an investigation from a one-on-one chat, group chat, channel or meeting chat. When the agent wants to take an action that needs sign-off, it posts an approval card into the conversation. Administrators decide which shared conversations the agent may join. Existing tool policies decide whether each action is allowed, needs approval or is denied. Access to Azure resources still runs through a managed identity and role-based access control.
The Teams update was one of several in Microsoft's Ignite agentic operations post, published on 16 November and updated on 18 November. SRE Agent also gained customer-managed keys, Entra Agent ID, Agent 365 integration and cross-agent communication, all in public preview. The Azure Resource Manager MCP Server could already read more than 3,500 resource types. Agents can now use the same server to create, manage and update resources, and Azure Policy evaluates those requests the same way it evaluates a change made in the portal. Heise's Ignite coverage points out that the new SRE Agent features are a public preview for now.
Microsoft's own usage figures are large. The company says it runs more than 11,000 SRE Agent instances internally, seven times as many as when the product reached general availability in March. It also says that over the last 90 days an average of 10% of its incidents were handled end to end with no human involved, and that remediation time fell from over 100 minutes to about four. Microsoft reported all of these numbers itself and nobody has verified them independently. A hyperscaler's operations team also has far more runbooks and telemetry to work with than a 200-person NZ firm.
Where the agent meets your service desk
Microsoft's documentation shows how SRE Agent connects to an ITSM tool, and service managers should read it before anyone switches the agent on. According to the incident platforms page, the agent accepts incidents from Azure Monitor, PagerDuty or ServiceNow. Only one of those platforms can be connected at once. With ServiceNow connected, the agent can read incident details, post discussion entries, acknowledge incidents and resolve them. The agent can close a ticket, then, but nothing in the documentation says it raises a change record or updates the CMDB.
The setup wizard can also create a quickstart response plan, and service managers should look at that plan closely. For Azure Monitor the quickstart plan picks up Sev0 to Sev2 alerts, and for PagerDuty it picks up P1 incidents. In both cases the plan runs in fully autonomous mode. Microsoft warns that leaving the quickstart plan in place next to custom plans can send incidents to the wrong agent or process them twice. Your most serious incidents are the ones the agent would handle first and without approval.
The mitigations documentation explains what the agent can actually do. SRE Agent can run any Azure CLI command its managed identity has permission for. Microsoft's own examples include restarting web apps, resizing VMs and creating role assignments. Microsoft has blocked a few things outright: delete and remove commands, all Key Vault commands, and changes to resources with ReadOnly locks. Every command the agent runs is logged to Application Insights as a custom event. For human review and change management, the documentation suggests having the agent create a GitHub issue or an Azure DevOps work item.
Put together, the documentation describes an agent that can make production changes and then resolve the incident. The evidence of those changes sits in Application Insights and possibly a Teams thread, while your change calendar and CMDB stay unchanged. An auditor or a post-incident review would struggle to rebuild what happened. Restarting an App Service at 3am under an approved runbook is a reasonable standard change. Granting a role assignment or resizing a production VM is a different kind of change, and SRE Agent's run modes apply to a whole response plan, with no setting for individual commands.
What this means for NZ organisations
Plenty of NZ mid-sized organisations run their Azure estate through a managed service provider. For those firms the first risk is a provider enabling SRE Agent on a customer subscription as a productivity tool, giving the managed identity Contributor rights on a resource group and leaving the quickstart plan running. Microsoft's own advice is to start at single-resource scope, use resource group scope for production, and keep subscription scope for development and testing. Ask your provider which scope they've granted.
Cost is the second question. Since 15 April 2026, SRE Agent has been billed in Azure Agent Units for every million tokens used in active flows. An agent that investigates every Sev2 alert across dozens of services will use tokens all day, and Microsoft hasn't published typical monthly costs for a mid-sized estate. Run a short pilot on one workload before you commit, and put a budget alert in place for the pilot.
The Ignite announcement also doesn't say which regions the new preview features run in, or where investigation data and agent memory are processed. If your incident data includes personal information, check that against your privacy obligations before you connect production workloads.
None of this is a reason to ignore the tool. Investigating a recurring memory alert or a failed deployment is exactly the repetitive work an agent can take off a small operations team. Run the agent inside your change process from day one, so nobody has to rebuild controls around it after something goes wrong.
Practical next steps
- Delete the quickstart response plan if it's running, and build response plans in Review mode with severity filters you've chosen yourselves.
- Agree which agent actions count as standard changes under your ITIL 4 change enablement practice. Restarts and scale-outs on named services are reasonable candidates. Role assignments and SKU changes belong in normal change.
- Scope the agent's managed identity to single resources or one resource group, and record who owns the identity in your CMDB.
- Make sure every autonomous fix leaves a record your service desk can see. You could export the Application Insights events into ServiceNow, or have the agent create a linked work item.
- Limit the Teams channels where the agent can post approval cards, and record approvers the same way you'd record a change advisory decision.
- Ask your MSP in writing whether they've enabled SRE Agent, the ARM MCP Server write tools or Copilot agents on your subscriptions.
- Set an Azure budget alert for Agent Unit consumption before any pilot.
If you're working out how much autonomy to give AI agents across your IT operations, our AI consultancy team can help you set those rules before an agent starts resolving tickets.
Sources
- AZ580753: Public Preview: Microsoft Teams integration for Azure SRE Agent
- Advancing agentic operations on Azure at Ignite 2026 (Microsoft Tech Community)
- Microsoft Ignite 2026: Azure gets more agentic as Windows moves to the cloud (heise)
- Announcing general availability for the Azure SRE Agent (Microsoft Tech Community)
- Incident platforms in Azure SRE Agent (Microsoft Learn)
- Execute mitigations in Azure SRE Agent (Microsoft Learn)
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help