Datadog Releases Scout AI Model for SRE Production Problem Investigations

Datadog Introduces Scout AI Model

Datadog has launched Scout, an artificial intelligence model specifically trained to investigate production problems. Scout is integrated into Datadog's existing AI SRE product, Bits Investigation, where it processes alerts and telemetry data to suggest potential causes for issues. The model is currently in a closed preview with Datadog engineers and a limited number of customers.

Cost-Effective Investigations

A primary focus for Scout's development was reducing the cost associated with AI-driven investigations. General-purpose AI models can perform similar tasks, but their inference costs become prohibitive when applied to every alert. Scout is designed to offer a more economical solution, enabling Bits to initiate investigations automatically as a 'first responder' before human engineers intervene.

Datadog's internal AlertEval benchmark indicates that Scout medium achieves performance comparable to GPT-6.1 Sol high and Claude Opus 5.5 high, but with an estimated inference cost that is three to eight times lower. At its xhigh setting, Scout matched Claude Fable 5.1 xhigh at one-quarter of the cost, based on relative pricing to a Scout medium investigation.

Training and Evaluation

Scout is built upon GLM-5.3 Flash and was post-trained using reinforcement learning. Its training involved incident simulations created from patterns observed in historical cases, which included an alert, queryable telemetry, and a known cause, without using customer data. Scout practices with tools that mimic those in Bits, such as log, metric, and trace searches, as well as code, deployment, and feature flag inspections.

The AlertEval benchmark uses real alerts, archived telemetry, and expert explanations to evaluate Scout's performance. An LLM judge compares Scout's findings against expert accounts, rewarding accurate identification of immediate and underlying causes while penalizing contradictions or unsupported claims. The resulting score, ranging from zero to 100, is not a direct percentage of incidents solved.