When a production system goes down, every minute costs money. An engineer gets paged, opens their laptop, starts digging through dashboards and logs, forms a hypothesis, tests it, rules it out, and tries another angle while everyone waits. That cycle is exhausting. Datadog just built a model trained to handle the first part automatically, before a human is ever paged.
Scout 1.0 is Datadog’s first AI model built specifically for investigating production incidents. It powers Bits Investigation, the company’s AI-assisted SRE tool. SRE stands for Site Reliability Engineering, a discipline focused on keeping production systems healthy and responding when they fail. Scout is designed to run as an always-on first responder, analyzing alerts and gathering evidence before an engineer looks at the screen.
The key differentiator isn’t raw accuracy alone. Scout matches the investigation quality of GPT-6.1 Sol and Claude Opus 5.5 at 3 to 8 times lower inference cost. Inference cost is what you pay to run a model per call. Real incident investigation involves dozens of calls. Scout forms hypotheses, queries telemetry (the continuous stream of performance data flowing from your systems), checks logs, and tests competing explanations step by step. Lower cost per call means Datadog can run Scout on more alerts, earlier in an incident’s lifecycle, and at greater depth.
Datadog trained Scout using reinforcement learning, a technique where a model improves through practice and feedback. The training data came from thousands of historical incidents with known root causes, combined with real telemetry from those events. No customer data was used. Scout learned to use the same tools Bits Investigation provides, querying metrics, correlating events across services, and identifying what broke. It was built for this specific job, not adapted from a general-purpose model after the fact.
That distinction matters more than it sounds. General-purpose models are designed to do everything adequately. Scout was trained to do one thing well: look at a production alert and find the cause. In Datadog’s AlertEval benchmark, Scout at its highest effort level outperforms GPT-6 Astra and matches Claude Fable 5.1, while costing far less to run.
The practical payoff for engineering teams is direct. When an alert fires, Scout runs an investigation before anyone is paged. The on-call engineer gets context, a likely cause, and proposed recovery steps rather than a blank screen and a racing heart at 3 AM. Datadog is already working toward expanding Scout into autonomous remediation, where it doesn’t just diagnose but fixes.
Scout 1.0 is in closed preview now. General availability is coming.
Want to explore how AI-powered operations could benefit your business? Let’s talk.