Operations | Monitoring | ITSM | DevOps | Cloud

When evaluating AI models for observability, generating a plausible query or explanation isn't enough. Teams need to know whether an agent's conclusions are supported by telemetry, whether its changes to dashboards actually work, and whether it can deliver reliable results consistently. This is why we built o11y-bench, Grafana Labs' open benchmark for evaluating AI models on observability and SRE tasks. And today, we're excited to share two updates.

Agents have quickly become part of the tech landscape as organizations embrace AI across their businesses. But embracing agents also means embracing the monitoring best practices that have long been part of engineers' workflows: tracking performance, catching errors, and responding to incidents But it doesn't stop there. Agents also introduce new challenges.

You're investigating checkout errors when someone asks who's coordinating the incident. Support needs an update, and another responder is looking for the runbook. You stop what you're doing to get everyone pointed in the same direction. They're all reasonable requests, but each one takes your attention away from understanding the problem. Yes, your team has a process for handling incidents, but carrying it out is another job for someone who's already busy.

Most teams have telemetry from across their stack, but connecting those signals to understand what's happening and how it affects users isn't always easy. Critical context about service dependencies, ownership, and past incidents often lives across tools, runbooks, and conversations between team members. That’s why, today at ObservabilityCON 2026, we’re announcing new capabilities that advance full-stack observability across Grafana Cloud.

In the AI era, where telemetry is generated by every model call, tool invocation, and agent step, organizations need more help than ever when it comes to navigating the economics of observability. For a lot of teams, the default response is a "data diet" of shorter retention, which leaves services or environments out of collection, or basic sampling, which is more guesswork than strategy. Either way, you end up cutting to control the bill, and you risk losing the very data you need during an incident.