Introducing the Evaluation Engine: Build High-Accuracy Evaluators and Guardrails Fitted to Your Agents
Splunk, Tuesday, November 10th, 2026
Splunk introduces the Evaluation Engine in Agent Observability for building custom agent evaluators and guardrails from traces and feedback.
Splunk Agent Observability now includes the Evaluation Engine, which turns traces and expert feedback into custom evaluators and runtime guardrails that follow an organization's own policies.
Its evaluation tuner generates training data and tunes specialized small models, which Splunk says can be done in an afternoon without a data science team.
Splunk reports that its Eval-G1 models match the most accurate frontier LLM judges on PII and toxicity, with the PII evaluator catching 99.4% of leaked personal data at 0.958 F1.
Each verdict takes about a third of a second, up to 14 times faster than frontier LLM judges and at up to 98% lower cost, and the feature is available now to Splunk Agent Observability customers in US regions.