Batch Telemetry Agent

Episode recorded across three time zones - Singapore, US, and Sweden

Johan and Gustav welcome Prashant Verma (Microsoft Singapore ISD) and Hemanth Kumar (Microsoft Azure FastTrack) for a technical deep dive into the batch telemetry agent. Prashant's origin story reveals that the agent emerged from his Premier Field Engineer days, handling APAC escalations for slow-running batch jobs and throttling issues, and recognizing that customers needed self-service access to insights previously requiring support tickets.

The conversation positions the agent as a "team member" that democratizes root cause analysis, bridging the gap between technical Application Insights data and business users who lack Kusto query skills. Hemant emphasizes the transformation: rather than context-switching between Teams conversations and Application Insights dashboards, system admins can troubleshoot directly within Teams, where business users are already reporting issues.

Johan's thesis emerges: fewer people logging into Finance & Operations is actually better since it's not intuitive for non-daily users. The agent supports this by making troubleshooting accessible to IT staff unfamiliar with F&O's interface. Gustav tests the waters on autonomous monitoring - could agents proactively ping Teams when detecting anomalies? The answer: marry Application Insights anomaly detection with agentic analysis to create alerts that inform users of slow critical jobs and suggest remediation.

The MCP server integration (version 46 required) enables self-healing actions, such as restarting batch jobs, though both hosts and guests emphasize guardrails. Gustav suggests SharePoint lists indicating which jobs are safe to auto-restart; testing reveals Claude 4.5 best follows these instructions without hallucination compared to GPT models.

A revealing war story: a customer configured batch jobs requesting 300 threads when their environment only supported 160 total threads, strangling the entire system for a week until Prashant identified it through telemetry. The agent now surfaces these misconfigurations instantly.

Setup requires just five minutes with existing Application Insights (ISM telemetry no longer needed), plus 15 minutes for agent installation. Model selection matters: 4.1 suffices for simple queries, 5.0 for complex analysis, and Sonnet for query generation.