This talk explains two connected jobs: modernizing software and restoring service safely. It is a conceptual walkthrough and a proposed nonproduction exercise, not a report of a completed deployment.
Modernization changes the application over time. Incident response restores an acceptable operating state. The loops connect through evidence: a recurring incident can become an issue with logs, a reproduction, and a proposed durable fix. "ZeroOps" describes an ambition to reduce toil, not the disappearance of operational responsibility.
The model does not directly restart a service. A runtime invokes a tool under configured permissions and policies. Resource permissions answer what an identity can access. Approval and tool policy answer which actions need consent or are denied. Review mode gates Azure infrastructure actions; other connector actions need their own policies or hooks. Reference: https://learn.microsoft.com/en-us/azure/sre-agent/run-modes
This is a logical architecture, not a network or resource-containment diagram. Start with one nonproduction App Service application. Keep application code, infrastructure definitions, tests, and an operational runbook together in the repository. The approval gate represents human consent plus tool policies; the agent's identity separately needs the required resource permission. Repository Code Access and GitHub issue/workflow connectors are separate capabilities. References: https://learn.microsoft.com/en-us/azure/sre-agent/github-connector and https://learn.microsoft.com/en-us/azure/sre-agent/overview
An upgrade changes frameworks, dependencies, and project formats. A migration assesses readiness and plans supported changes for a target platform. The work can be done separately or together. Build success alone does not establish that business behavior is unchanged. References: https://learn.microsoft.com/en-us/dotnet/core/porting/github-copilot-upgrade/overview and https://learn.microsoft.com/en-us/azure/developer/github-copilot-app-modernization/overview
Connect telemetry and an availability alert to SRE Agent. Set the response plan to Review before stopping a nonproduction app. Inspect the diagnosis and proposed start action; approve only the action scoped to that application. Check the endpoint and a representative user journey. The pass/fail branches are the outcome; command success alone is not recovery. The example's consecutive checks and one-attempt limit are proposed lab rules, not measured results or production objectives. Reference: https://learn.microsoft.com/en-us/azure/sre-agent/run-modes
These controls solve different problems. An approval does not grant resource permission. Review gates Azure infrastructure actions, not every connector write; other tools require their own policies or hooks. Use scoped identities, branch and environment protections, audit trails, bounded actions, and a named escalation owner. References: https://learn.microsoft.com/en-us/azure/sre-agent/run-modes, https://learn.microsoft.com/en-us/azure/sre-agent/github-connector, and https://docs.github.com/en/actions/how-tos/secure-your-work/security-harden-deployments/oidc-in-azure
Documentation checked November 12, 2026.