Agentic AI
Brain: Microsoft's AI outage detector now makes autonomous repair decisions across Azure
Azure's internal AI system monitors cloud health in real time, declares outages, halts flawed deployments, and increasingly uses agents to fix incidents automatically.
The New Stack
Microsoft has unveiled Brain, an internal AI operations platform that watches Azure's health continuously and now acts on what it detects — flagging outages, stopping problematic rollouts, and alerting customers before they discover problems themselves.
The system represents years of infrastructure work that predates the generative AI wave. Azure CTO Mark Russinovich, who detailed the project in a blog post, explained how Brain evolved from a foundational effort to build a real-time picture of how Azure's services relate to and depend on one another.
From dependency mapping to intelligent triage#
Brain sits atop Azure Resource Graph, which began as an internal attempt to model relationships between Azure resources. Early on, the team realized they could trace service failures along these dependency chains — if Service A goes unhealthy and Services B, C, and D all rely on it, those downstream failures follow predictably.
A parallel problem pushed the work forward: Azure's own health metrics often showed everything working fine while customers experienced real outages. The disconnect came from how measurements were aggregated or what telemetry was actually collected. To solve this, Microsoft standardized on service level indicators (SLIs) — a shared measurement schema that services across Azure emit through common libraries.
Building consensus across hundreds of services to adopt the same measurement standard took several years of data engineering and coordination.
Three signal streams fuel decision-making#
Today, Brain ingests signals from three sources: the standardized SLIs from service teams, domain-specific monitors and deployment telemetry that teams build and register themselves, and third-party indicators like customer support tickets and social media mentions of Azure issues.
From any of these signals — whether they concern a service, a region, a deployment cluster, or a specific customer's resources — Brain produces four outputs: health status, severity assessment, a list of affected parties, and the reasoning behind its conclusion.
Those conclusions trigger actions automatically. Brain declares outages based on the number of affected customers (blast radius), scopes notifications only to impacted subscriptions and regions, routes incidents to the correct service team, and sends signals to deployment gates that pause rollouts causing problems.
