Datadog Scout 1.0: RL Post-Training Delivers a Cheaper AI SRE Model

Datadog Scout 1.0 is a specialized model for investigating production incidents. Datadog built it by post-training the open-weight GLM-5.3 Flash with reinforcement learning. On Datadog’s own benchmark, it matches frontier models at a fraction of the inference cost.

Scout powers Bits Investigation, Datadog’s AI site reliability engineering (SRE) agent. Datadog trained it in simulations built from real historical incidents and alert telemetry with known root causes, and the company says no customer data was used. Rewards favor finding the correct cause and backing it with evidence. Training ran on asynchronous RL infrastructure built on SkyRL with a modified GRPO.

On AlertEval, Datadog’s internal benchmark annotated by human experts, Scout 1.0 medium matches GPT-6.1 Sol and Claude Opus 5.5 at 3–8× lower estimated cost. Scout 1.0 xhigh matches Claude Fable 5.1 xhigh at 4× lower cost.

The takeaway for post-training is that RL on domain-specific environments can turn a mid-size open model into a frontier-level specialist. That lets teams run investigations far more often and earlier in an incident.