Databricks Deploys AI SRE to Automate Incident Debugging

Databricks has integrated an internal artificial intelligence tool called AI SRE to streamline how its engineering teams manage and troubleshoot system incidents. The AI assistant helps the company operate hundreds of microservices deployed across more than 1,500 Kubernetes clusters, spanning over 70 regions and three major cloud providers. Currently, the platform supports more than 150 teams and over 250 weekly active users, executing more than 2,000 daily investigations and saving several hours of debugging time per incident.

Before implementing the tool, Databricks found that context assembly consumed 60% to 80% of an on-call engineer's investigation time. To address this, AI SRE initiates automatic triage the moment an incident is triggered. It simultaneously runs platform health checks, performs service-level analysis on logs and metrics, and executes team-specific agentic runbooks. This parallel processing allows the system to correlate signals and present a comprehensive diagnostic summary by the time an engineer opens their laptop.

The architecture of AI SRE relies on a layered platform consisting of data primitives, a unified API layer, a core engine for orchestration, and an application layer. To ensure reliability in high-pressure production environments, Databricks designed the agent to prioritize structured, deterministic checks before applying large language model reasoning. Furthermore, the system emphasizes transparency by linking every diagnostic recommendation directly to verifiable raw evidence, such as specific metrics or log lines, allowing engineers to audit the AI's conclusions.

For site reliability practitioners, this development shifts the on-call experience from manual data gathering to evidence-backed verification. Instead of navigating multiple disconnected monitoring tools, engineers receive pre-filtered context in a single view. Databricks plans to expand the agent's capabilities in the future by moving into guided mitigation and utilizing cross-incident learning to identify systemic reliability gaps before they trigger alerts.