Datadog launches Scout, a troubleshooting model for AI SRE

On November 12, 2026, Datadog released Scout, a lightweight specialized model for investigating production issues. Scout runs inside Bits Investigation, Datadog's AI SRE product, where it takes in alerts, looks at telemetry data, and offers hypotheses about root causes. It is currently in closed preview, being used by Datadog engineers and a small group of customers.

Most of the announcement focuses on cost. Investigating a failure can take many model calls, because Bits tries out queries, analyzes the results, and decides what to check next. General-purpose models can do this kind of work, but Datadog says the high cost of calling them makes it impractical to run one on every alert. The team wants Bits to start investigating as soon as an alert fires, while engineers are still on their way in. "Our goal is to make Bits a first responder on call around the clock, gathering clues and pinpointing likely causes before engineers step in," the announcement says. Datadog adds that Scout is optimized for low latency and high-frequency alert responses.

On Datadog's AlertEval benchmark, the announcement reports that Scout medium scores close to GPT-6.1 Sol high and Claude Opus 5.5 high, while costing one third to one eighth as much per investigation. At the xhigh tier, Scout's results are said to match Claude Fable 5.1 xhigh at one quarter of the cost. Scout is built on GLM-5.3 Flash and was post-trained with reinforcement learning on simulated incidents. The announcement was authored by Othmane Abou-Amal, Ben Cohen, Afshin Rostamizadeh, Ameet Talwalkar, Omri Sass, and Daniel Shan.