Grafana applies SLOs and error budgets to AI agent behaviour

Grafana Labs outlined an approach that turns AI agent behaviour — hallucination, task fulfilment, toxicity, cost — into evaluations and SLIs measured against SLOs. In one example, 95% of conversations must be judged fulfilled over 30 days, leaving a 5% error budget for experimentation and drift detection.

• Agent behaviour is evaluated and converted into an SLI measured against an SLO
• Example SLO: 95% of conversations fulfilled over 30 days, 5% error budget
• Tracking error-budget burn surfaces gradual quality degradation
• Quality is multidimensional: groundedness, toxicity, cost and privacy measured separately