Google SRE AI Insights: historical incidents as investigation context

Corrected source check October 24, 2026: Stevan Malesevic and Christopher Heiser, “AI in SRE: Where and how Google is deploying agentic AI to improve operations”, May 28, 2026. Google SRE reports building AI Insights to extract information from known incidents, using embeddings and vector databases so investigation agents can consult that history. It also reports agents that maintain playbooks after incident use, and an orchestration layer for handoffs, incident summaries and communications. The post offers an architectural account of embedding old lessons in a search path, not a dated public trace of an earlier false diagnosis being corrected and applied to a second incident. It does not publish incident identifiers, a skill diff, a second agent’s read trace, a verified resolution or human minutes saved.

The primary article calls for explicit agent identity and roles, monitoring of AI actions, privacy and security requirements, production-state mutation controls and fallback plans; these are design principles, not proof of all controls’ enforcement in a given case. It also says some issues can be handled autonomously, so do not replace its mixed authority model with a claim that every action needs human approval. Contrast Sridharan's specific control overturning two agents' TLS diagnosis and Kiro's skill promoted from a per-cell incident. Google's post cannot establish correction→retrieval→action→field outcome; look for case-level traces rather than treating continuous indexing as proof of useful learning.