Enable Evaluators
Reading traces one at a time is great for an incident, but it doesn’t scale. Evaluators automatically evaluate your agent’s interactions so problems (hallucinations, tool-selection errors, prompt injections) surface on their own.
Persona
As Careful Health Provider’s AI engineer, you want to know the moment the assistant
produces an ungrounded medical answer, not when it trends on social media. Out-of-the-box
evaluators give you that early-warning system across every request.
Accurate evaluations. Splunk Agent Observability ships out-of-the-box evaluators
that evaluate agent, output, and RAG quality, so you can quickly spot problems, track improvements, and make your agentic applications work better for your users.
Evaluations that scale, powered by Luna
Using a large general-purpose LLM as a judge gets expensive and slow fast, so most teams
sample a fraction of traffic and leave the rest unscored. Splunk Agent Observability uses
Luna Small Language Models, purpose-built and fine-tuned for evaluation, to score quality
at a fraction of the cost and latency. That means you can evaluate all of your production
traffic (not a 5–10% sample) and even run evaluations fast enough for real-time guardrails.
Continue to enable evaluators on your agent stream and review the scores.