Grafana Labs proposes SLOs and error budgets for AI agents

On November 1, 2026, InfoQ reported on a post by Grafana Labs proposing to measure how well AI agents actually perform by combining evaluations with service level objectives (SLO) and error budgets. The approach starts by evaluating a single agent interaction against specific behavioral criteria: whether the response is grounded in the context provided, whether it completes the user's request, whether it contains toxic content, whether it leaks personal information, and whether it is affected by prompt injection. Those checks produce SLIs, or service level indicators. The judge can be another language model, but deterministic checks such as regular expressions work when the relevant behavior can be identified without an LLM. Grafana Labs argues that quality is multidimensional: an agent can be grounded yet expensive, helpful yet occasionally toxic, or accurate yet slow. Tracking those dimensions as separate metrics lets a team set targets for the behaviors its application actually cares about, rather than hiding important failures inside a single agent quality score.

Once a behavior can be measured, standard SRE concepts take over. Grafana Labs gives an example SLO requiring that within 30 days, 95% of conversations be judged to have completed the user's request. The remaining 5% becomes the error budget, the amount of unacceptable behavior the team explicitly decides it can tolerate in that window. While the budget remains healthy, a team can push changes such as prompt or model adjustments; once the budget is exhausted, the priority shifts to fixing quality. The company stresses that the value of this method lies less in the percentage itself than in the decision model it establishes, and that the hard part is not setting an SLO but deciding what target is actually acceptable. As of the report, this remains a proposal from a single Grafana Labs post, with no mention of outside teams having adopted it.