Datadog has launched Scout, an artificial intelligence model designed to diagnose production failures at a fraction of the expense of general-purpose systems. When tested at its highest performance tier, Scout matched Claude Fable 5.1's accuracy while costing just one quarter as much, according to an announcement from the monitoring company. The model operates within Bits Investigation, Datadog's AI-powered site reliability engineering tool, where it can receive an alert, examine telemetry data, and recommend a probable cause before a human engineer begins troubleshooting.
The company's internal benchmark, AlertEval, showed Scout's medium configuration scoring similarly to GPT-6.1 Sol high and Claude Opus 5.5 high, with estimated inference expenses running three to eight times lower. On SREGym, a publicly available collection of 90 system faults, Scout resolved 71 incidents at medium effort, 79 at high, and 80 at the highest setting, while GPT-6.1 Sol's most resource-intensive configuration completed 82. AlertEval relies on authentic alerts, stored telemetry, and explanations authored by Datadog's own engineers, with an LLM judge awarding points for matches at both immediate and root cause levels while penalizing contradictions and unsupported assertions. The benchmark produces scores between zero and 100, though Datadog cautions this isn't equivalent to a percentage of successfully investigated incidents. Scout is currently available to Datadog staff and a limited set of customers in a closed preview, with no general release date announced.
The authors state their objective is for Bits to function as "an always-on first responder, gathering evidence and identifying likely causes before an engineer steps in." Datadog built Scout through post-training GLM-5.3 Flash using reinforcement learning with incident simulations derived from patterns in historical cases, according to the technical report. The simulations include an alert, queryable telemetry, and a known cause, but contain no customer data, the company says. During practice sessions, Scout uses tools mirrored from those in Bits, including searches of logs, metrics, traces, code, deployments, and feature flags, with training that "rewards finding the cause and supporting its conclusions with evidence."
Cost drove much of the engineering work behind Scout. An investigation might demand numerous model calls as Bits attempts a query, reviews the output, and determines the next step to take. General-purpose models can handle this workflow, but Datadog says the expense makes running them against every alert impractical. The technical report describes how an early reward structure based solely on final answers encouraged Scout to make fewer claims and halt investigations prematurely, reducing penalties without improving its diagnostic ability. Engineers adjusted the reward to account for correctness, exploration, and supporting evidence, and removed tools from certain training runs or shortened available time to prevent the model from memorizing a single path through incidents. Independent verification would strengthen the benchmark results, since Datadog conducted all tests itself, including runs for newer comparison models.
Scout enters a market where observability vendors are deciding how much of an incident response an agent should manage. AWS DevOps Agent combines telemetry from multiple monitoring products with source code, deployment records, and runbooks, while Honeycomb's Canvas Agent runs queries and tests hypotheses when an alert fires, leaving evidence visible for human review. Splunk and New Relic have extended into the repair phase, with Splunk's connected Claude agent capable of preparing code changes and opening pull requests, and New Relic's Autopilot suggesting remedial actions, though both retain approval steps before production changes. Datadog plans to expand its simulations to cover more failure types and eventually enable Bits to perform repairs, but the current preview stops at investigation, diagnosis, and proposed recovery actions, with engineers retaining responsibility for changes. The independent AI news account AGTP noted that AlertEval hasn't undergone independent testing and requested an audit trail showing which evidence Scout considered and rejected, echoing a requirement other vendors address through visible investigation records. For organizations weighing the tradeoff between autonomy and oversight, the question isn't whether AI can investigate incidents, but whether teams will trust a model's reasoning when the diagnosis arrives faster than they can validate its logic. As vendors race to automate more of the response cycle, the gap between a correct answer and an explainable one may determine which approaches engineers actually adopt under pressure.