Grafana is adapting service-level objectives and error budgets to AI-agent behavior, giving engineering teams a way to track failures that ordinary uptime metrics miss. An agent can respond quickly with no technical error while still hallucinating, ignoring a request or exposing personal information.
The proposed workflow evaluates individual interactions for separate qualities such as groundedness, task completion, toxicity, prompt-injection resistance and cost. Those results become service-level indicators. A team could then require, for example, that 95% of conversations be judged fulfilled over 30 days, leaving a 5% error budget for unacceptable outcomes.
Keeping the dimensions separate avoids hiding a serious weakness inside one aggregate quality score. Error-budget burn can also reveal a slow decline after a prompt, model, retrieval system or user population changes. When the budget is healthy, teams can experiment; when it is nearly exhausted, remediation takes priority.
The method does not solve the hardest measurement problem. Language-model judges can misclassify answers, and poorly designed tests may reward responses that satisfy a formal rule without helping the user. Grafana therefore points to representative datasets, deterministic checks where possible, human review and continued validation of evaluators. The SLO creates an operational decision framework, but its reliability still depends on the evidence beneath it.