# Datadog Launches Scout AI Model for Automated SRE Incident Investigations Specialized model reduces inference costs while maintaining performance parity with general-purpose LLMs in SRE tasks. - **Source:** [InfoQ](https://brifing.si/source/68ac18af-a4f3-4f7c-9c2c-44d4af8f89dd) - **Published:** 2026-10-29 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/3c12d25a-7cfc-4cbc-9aff-72df07e9cef1 ## Specialized Model for SRE Datadog has introduced Scout, a specialized AI model designed to investigate production incidents within its Bits Investigation platform. Currently in closed preview, the model analyzes telemetry data to identify root causes. By focusing on site reliability engineering workflows, the company aims to provide faster diagnostic insights during critical system alerts. ## Optimizing Inference Costs The primary driver for developing Scout is the high cost associated with using general-purpose models for frequent alert investigations. Datadog reports that Scout medium achieves performance scores comparable to GPT-6.1 Sol and Claude Opus 5.5, while reducing inference costs by three to eight times depending on the specific configuration used. ## Training via Incident Simulations Scout is built upon the GLM-5.3 Flash architecture and refined through reinforcement learning. Datadog utilized historical incident patterns to create simulations that include alerts, queryable telemetry, and known causes. These simulations allow the model to practice using tools like log searches and feature flag inspections without accessing actual customer data. ## Refining the Reward Mechanism During training, engineers discovered that rewarding only the final answer encouraged the model to stop investigations prematurely. To improve accuracy, the team adjusted the reward function to prioritize exploration and supporting evidence. They also introduced constraints, such as removing tools or shortening time limits, to prevent the model from learning fixed paths. ## Performance Benchmarking Results Datadog evaluated Scout using its internal AlertEval benchmark and the public SREGym dataset. On SREGym, Scout resolved 80 out of 90 faults at the xhigh setting, compared to 82 resolved by GPT-6.1 Sol. While these results are promising, the company notes that AlertEval scores do not represent a direct percentage of incidents solved. ## Competitive Observability Landscape The industry is currently debating how much autonomy AI agents should have during incidents. While AWS focuses on integrating telemetry across tools, Honeycomb emphasizes transparency through its Canvas Agent, which displays the investigation process. Splunk and New Relic have moved further toward automated repair, though they maintain human approval for production changes. ## Transparency and Auditability Concerns Independent observers have raised questions regarding the lack of third-party validation for Datadog's AlertEval benchmark. Critics emphasize the need for an audit trail that shows exactly which evidence the model considered or discarded. Similar to Honeycomb’s visible evidence chains, engineers require clear visibility into how an AI model reaches its diagnostic conclusions. ## Future Roadmap and Scope Datadog has not announced a general availability date for Scout. Future development plans include expanding the types of failures covered in simulations and eventually enabling Bits to execute repairs. For now, the model remains limited to investigation and recovery proposals, ensuring that human engineers retain final responsibility for all production changes. ## Sources covering this story - https://www.infoq.com/news/2026/10/datadog-scout-sre-model/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._