Datadog launches Scout 1.0 for production issue checks

Datadog has introduced Scout 1.0, an AI model for investigating production issues. It is the first in-house system the company has built specifically for site reliability engineering investigations.

Scout 1.0 powers Bits Investigation, Datadog's AI site reliability engineer tool. It is designed to examine production alerts before a human engineer begins work. Datadog trained the model to reduce inference costs so investigations can run more often and earlier in an incident's lifecycle.

The move reflects a broader push by software suppliers to build smaller, task-specific AI systems rather than rely solely on general-purpose models. For Datadog, the focus is a routine but resource-intensive part of operations work: checking telemetry, testing possible explanations and narrowing likely root causes when systems fail.

Each investigation involves repeated model calls as the system forms hypotheses, queries telemetry and assesses whether the evidence supports a conclusion. Lower inference costs give teams more scope to run those investigations across their environments and apply the system to more alerts.

Cost focus

Datadog said Scout 1.0 medium matched GPT-6.1 Sol high and Claude Opus 5.5 high on its target scenarios at an estimated inference cost three to eight times lower. It also said the model performed comparably to Fable medium and Astra medium at 12 times and 19 times lower cost, respectively.

With additional test-time scaling, Scout 1.0 xhigh matched Fable xhigh at four times lower cost, according to Datadog. The company said that lower cost makes it practical to devote more computation to harder or higher-impact incidents without making the overall process too expensive to deploy widely.

That cost argument is central to the product's design. Production incident analysis often requires repeated rounds of questions and checks, which can quickly drive up compute bills when teams rely on large frontier models. By building a specialised model tuned for this workflow, Datadog is trying to make automated investigation common enough to become part of normal incident response rather than a tool reserved for exceptional cases.

Training method

Scout 1.0 was built on GLM-5.3 Flash and post-trained using reinforcement learning in simulated investigations. Datadog used historical incidents and telemetry from real alerts to create simulations with known causes, while excluding customer data from those training simulations.

The company also built its own distributed reinforcement learning infrastructure using open-source tools. Scout was trained to use tools modelled on Bits Investigation so it could gather evidence, test competing explanations and support conclusions with data.

To assess whether those skills carry over to real incidents, Datadog created an internal evaluation framework called AlertEval. Engineers curated thousands of evaluation points through its incident-response workflow, combining real alerts, archived telemetry and root-cause explanations written by engineers.

Scout was also tested on SREGym, a third-party benchmark for site reliability engineering tasks. Datadog said Scout 1.0 medium improved on the underlying GLM-5.3 Flash base model and surpassed GLM-5.3 at more than four times lower cost. It added that Scout 1.0 high was competitive with leading models on that benchmark while maintaining at least double the cost savings against all competitors and 10 times the savings against Astra- and Fable-class models.

Early use

Datadog engineering teams are already using Scout to investigate production alerts, alongside a small group of customers in a closed preview. The company is therefore testing the model both in its own operations and in external environments ahead of a wider release.

That internal use matters because Datadog sells observability and monitoring software to engineering teams managing large, often complex cloud systems. A model that can handle the first stage of incident investigation could reduce the manual triage required from on-call staff, especially when alerts are frequent or arrive outside normal working hours.

Datadog's longer-term aim is for Bits to act as an always-on first responder for every alert by gathering evidence and identifying likely causes before an engineer steps in. In more complex or higher-stakes incidents, Scout 1.0 xhigh is intended to use more computation per incident to improve investigation quality.

The company is also developing the model and the investigation product together, training Scout for the tools Bits uses and adjusting those tools based on what it learns from investigations. That suggests Datadog sees the model less as a standalone release and more as part of a tightly controlled product workflow within its software stack.

Datadog said preview customers are already using Bits Investigation powered by Scout on their production alerts.