Latest in AI
Azure SRE Agent in Teams: Check Chat Access and Azure Permissions Separately
A rollout checklist for Azure SRE Agent's Teams preview, covering chat authorization, action approvals, managed identity and cross-team collaboration.
Microsoft's November 18 Azure SRE Agent update brings Teams integration into public preview. Putting an operations agent into an incident chat makes it easier to ask for help. It also makes a permissions mistake easier to overlook: being present in a conversation is not the same as being authorized to change infrastructure.
For Azure SRE Agent's new Teams integration, I would review three things separately: who can talk to it, which actions need approval and which resources its execution identity can reach. Treating those as one permission setting is the wrong starting point.
Start with the actual preview boundary
Microsoft's November 18 update describes native Teams integration, Agent-to-Agent collaboration and Entra Agent ID as public previews. Native Slack integration is in private preview.
An administrator grants access per chat and chooses who can interact with the agent. Action policy separately determines what is allowed, approval-gated or blocked. Entra Agent ID supplies a governed identity with an accountable sponsor; it complements rather than replaces the managed identity used for Azure access. Read the enterprise rollout announcement.
My starting scope would be one service, one incident chat and a named administrator. Prove that boundary before inviting multiple teams.
Inspect the real role assignments
Microsoft's permissions documentation says the agent uses a user-assigned managed identity. Connecting Teams or creating an Entra agent identity does not copy or replace that identity's Azure permissions.
There is a detail worth checking carefully: the documented preconfigured roles include Monitoring Contributor at subscription scope, even alongside the Reader permission level. That role supports alert lifecycle and monitoring changes. Do not interpret a “Reader” label as proof that every possible operation is read-only.
My recommendation is to inspect the assigned roles and their scopes directly. Record what the identity can change, not just the permission level selected during setup.
Design the approval path before an incident
The same documentation describes on-behalf-of authorization when the managed identity lacks access. Only SRE Agent Administrators using work or school Entra accounts can authorize those requests; ordinary users and personal Microsoft accounts cannot. Check the authorization rules.
For a pilot, I would make a short action register. Separate investigation, reversible operational changes and destructive or high-blast-radius operations. Name an approver and an escalation route for each permitted change.
Then rehearse what happens when that approver is unavailable. The agent should leave a useful investigation record, not create pressure to grant permanent broader access just to clear an incident queue.
Keep cross-team collaboration scoped
Microsoft says A2A collaboration uses configured connections and approved callers identified through Entra. A request to another team's agent does not give the requesting agent that team's infrastructure access. See the collaboration model.
I would agree on the information each team can share before connecting agents. Incident summaries can contain customer identifiers, internal topology or sensitive logs even when no resource modification occurs.
Use a small request contract: the affected service, time window, observed symptom and evidence needed. Avoid forwarding an entire conversation merely because it is convenient.
Test denials as well as successful requests
My proposed acceptance check has four scenarios: an approved user requests an investigation, an unapproved participant tries the same request, an operator requests an approval-gated change and a request targets an out-of-scope resource.
For each, retain the observed authorization decision and any resulting action. These are suggested tests, not results from a deployment I have operated.
Also check who can see the agent's responses in the chat. Permission to operate infrastructure and permission to read incident details are different design questions.
I would expand only after the team can explain why each request succeeded or failed. A helpful incident teammate should make your existing authority boundaries easier to follow, not harder to see.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help developer-tool teams make operational agents and their permission boundaries understandable through technical writing.