What is Azure SRE Agent (Microsoft)

Microsoft's AI agent embedded in Azure for autonomous incident response — acknowledges alerts from PagerDuty/ServiceNow/Azure Monitor, investigates via Azure Monitor/App Insights/Kusto, executes configurable Incident Response Plans with tunable autonomy, and learns across incidents via Session Insights.

Updated August 9, 2026 · §202608021430-6 ·

Azure SRE Agent is Microsoft’s AI-driven incident-response agent, built into Azure rather than deployed as a separate open-source project. Where What is HolmesGPT is a bring-your-own-infra CNCF Sandbox project you wire into whatever platform you already run, Azure SRE Agent is Azure-native by design — it starts investigating the moment an alert fires, without a human triggering it.


Architecture: alert-triggered agentic loop

Alert fires (PagerDuty / ServiceNow / Azure Monitor)


SRE Agent acknowledges the alert


Queries Azure Monitor, App Insights, Kusto, connected 3rd-party tools


Hypothesis-driven investigation (extended reasoning for complex incidents)


Root cause + proposed fix ──▶ executes or waits for approval, per configured autonomy

Incident Response Plans — routing and autonomy control

Response plans route an incoming incident to the right custom agent, at the right autonomy level, based on rules you define — filterable by severity, service, title, and type. This is the configuration surface: instead of one agent handling every incident the same way, different incident shapes get routed to different plans.

Run modes

ModeBehavior
AssistedInvestigates and proposes a fix; a human approves before anything executes
AutonomousFor scenarios trusted and well-tested enough to skip approval — the agent remediates directly

Learning loop: Session Insights

Unlike a static runbook, the agent generates a Session Insight after every incident — a structured summary of what happened, what worked, and what to improve — and that record feeds forward into how future incidents of the same shape get handled. The “learns and remembers” framing is the same design goal What is HolmesGPT and other agentic-RCA tools are converging on; the difference is where that memory lives and who controls it.

Where it sits next to HolmesGPT

ConcernAzure SRE AgentWhat is HolmesGPT
Vendor / opennessMicrosoft, Azure-native, closedRobusta.dev, Apache 2.0, CNCF Sandbox
Integration surfaceAzure-first: Azure Monitor, App Insights, Kusto, PagerDuty/ServiceNow/Azure Monitor incidents70+ toolsets across Kubernetes, VMs, cloud, databases, any infra
Autonomy controlPer Incident Response Plan — Assisted vs. Autonomous”Operator mode” — continuous background monitoring, opens fix PRs
MemorySession Insights per incidentNo equivalent persistent per-incident memory layer documented

Why it’s relevant here: the per-plan, configurable autonomy model (route by severity/service, choose assisted vs. autonomous per route) is a cleaner articulation of the same human-in-the-loop-by-design tension that Build an AI SRE Assistant (Part 00) works through with its own approval-gate design — worth comparing Microsoft’s routing-based autonomy control directly against that pattern, alongside HolmesGPT’s read-only-RBAC approach to the same problem, before extending that hands-on build further.

Local graph

Full graph →