Operations | Monitoring | ITSM | DevOps | Cloud

With SRE AI, Google plans to fully adopt AI and agentic technologies, leveraging AI as a force multiplier while also maintaining control.

A critical production alert wakes you up: p99 latency just hit 4 seconds. You drag yourself to a terminal, open five dashboards, start correlating log timestamps with trace IDs, dig through 47,000 log lines across eight services, and 90 minutes later, you finally find the culprit: an N+1 database query introduced in a deployment that shipped four minutes before the spike started. An Atatus AI SRE Agent would have identified that root cause and drafted a remediation plan in 28 seconds. Not approximation.

An error budget is the acceptable amount of unreliability permitted by your SLO over a defined time window. It is not a target. It is not a stretch goal. It is a hard ceiling that, when breached, should trigger a pre-agreed organizational response — feature freezes, postmortems, or infrastructure investment. The formula is blunt: Error Budget = 1 - SLO Target Error Budget (time) = (1 - SLO Target) × Window Duration For a 30-day window: That last number should make you uncomfortable.

Single agents are a useful starting point for SRE workflows. They are not where the architecture should end. The first version is simple enough: connect an LLM to a few tools, give it a system prompt, and point it at your infrastructure. It can summarize an alert, pull logs, answer questions, and draft a useful next step. Then the workflow gets real. You add GitHub for runbooks, Kubernetes for cluster state, PagerDuty for incident context, Prometheus for metrics, and Mezmo for telemetry.

Drop a JAR on the JVM. Get distributed tracing, RxJava context propagation, log-trace correlation, and Vert.x internal metrics. No code changes. No Maven dependency. Java 8–21. Inside the design of last9/vertx-opentelemetry v2.3.4. Prathamesh works as an evangelist at Last9, runs SRE stories - where SRE and DevOps folks share their stories, and maintains o11y.wiki - a glossary of all terms related to observability.

In a recent webinar, The Journey to Production AI, Andre Elizondo walked through what separates a working agent demo from an agent worth trusting on a 2 a.m. page. Live polls during the session put numbers behind a pattern most platform teams already feel. ‍ ‍ Most teams are early. The ones who are further along did not get there by shipping a flashier demo. They got there by treating production AI as a platform problem.

AI promise and AI capabilities often diverge, with developers often reporting much faster code production, but not enough change in how incidents are handled. When the rate of change is faster than ever, but the rate of recovery from incidents isn’t moving, developers wind up stuck in firefighting mode. And, when these systems fail, it’s costly. According to PagerDuty’s State of AI-First Operations, over a third of surveyed companies report losing $500K per hour of downtime.

Why ECS containers collapse under service.name = aws_ecs and how to fix it for both EC2 launch type and Fargate, including the resource-vs-log-record pitfall that quietly breaks log filtering. Prathamesh works as an evangelist at Last9, runs SRE stories - where SRE and DevOps folks share their stories, and maintains o11y.wiki - a glossary of all terms related to observability.

What actually works for Kubernetes monitoring at scale — not what looks good in a vendor demo with a five-pod cluster.