Agentic Monitoring in Production

Leonel says eBay’s business reliability platform combines site reliability practices with business metrics such as impressions, clicks, and sales. Only 3% of observed incidents and anomalies were problems that cost money, yet investigating the rest consumed time. More complex thresholds risked missing real issues, so the team sought to automate investigation practices.

Sujith says the team focused on one decision: whether an alert should be escalated, and what an on-call engineer needs to make that call. Leonel explains that they split the work into a diagnostic agent, which uses documentation to create a structured plan, and an enrichment agent, which checks live metrics and dimensions. The enriched alert goes to the engineer, who retains the escalation decision.

The system gives engineers troubleshooting steps and answers questions about whether a signal persists, where it occurs, which business segments are affected, and what changed. The diagnosis plan links to source documentation, while the enriched message lets engineers return to dashboards and verify the data. In an eBay Ads example, revenue and GMV fell alongside sales and transactions, with web users declining while mobile was unaffected.

Sujith says GPT-5 produced good summaries in staging, but at eBay’s scale it had high latency and timed out in pre-production. The team switched to GPT-4, which he says responded 50 times faster and was sufficient for the analysis. Their deployment supports 22 billion daily impressions and about $7 million in daily ad revenue.

The team learned to scope the decision first, split the work when the model could not take all the context, and choose a model for predictable production behavior. Leonel says investigation time fell from an average of 20 to 30 minutes to under five minutes, making investigations 70% faster. Engineers still validate the findings and decide whether to escalate.