Grafana extends SRE-style SLOs and error budgets to AI agent behaviour

The approach gives teams separate reliability signals for agents instead of one score, and can expose quality drift before failures become obvious.

Grafana Labs is extending SRE-style reliability tracking to AI agents by measuring behaviours like grounding, task completion, toxicity, prompt injection and cost with evaluations, SLOs and error budgets. The company argues this is better than a single quality score because different failures matter differently, and it can surface slow quality drift even when an agent still appears to work. Grafana’s example target is 95% of conversations judged fulfilled over a 30-day window, with the remaining 5% treated as tolerated error budget. The company says targets should be set from baseline evaluations and historical results, then tightened as systems mature, but warns that bad judges or weak evals can make the resulting budgets misleading.