6 СУДЖИТ Н. С. ЛИОНЕЛ КАНДИДО ДА СИЛВА

Only 3% of anomalies in eBay's business metrics point to a real revenue problem, yet each alert once sent engineers across dashboards, documentation, correlated metrics, and business dimensions for 20 to 30 minutes. Sujith N S and Leonel Candido da Silva show how eBay built an agentic monitoring workflow that completes the first investigation in under five minutes while keeping the escalation decision with the on-call engineer. Recorded at Agent Conf 2026 in Warsaw, the talk follows two production agents: a diagnosis agent that turns operational knowledge into an investigation plan, and an enrichment agent that analyzes current metrics, correlations, and affected segments. The speakers share the architecture, traceability controls, model selection lessons, human boundaries, and production results, including a 70% faster investigation process. In this talk: Why anomaly detection still leaves engineers with expensive manual triage Scoping the workflow around the decision to escalate an alert Turning operational documentation into a structured diagnosis plan Correlating live metrics across sites, pages, segments, and platforms Grounding agent output with source references and dashboard data Matching model capability to production latency requirements Keeping the escalation decision with the on-call engineer Reducing investigation time from 20 to 30 minutes to under five minutes Chapters: 00:00 Introduction 00:23 Speakers and eBay's business reliability platform 00:50 Anomaly detection versus investigation 03:02 Turning alert triage into an agentic workflow 03:33 How much should engineers trust an agent? 04:53 Investigating an 18% revenue drop 05:50 The manual investigation workflow 06:38 Defining the escalation decision 07:28 Context window limits 08:03 Diagnosis and enrichment agents 09:46 The end-to-end workflow 10:50 From a raw alert to an enriched notification 13:02 eBay Ads investigation example 15:02 Model choice and production latency 16:46 Building traceability into the workflow 17:28 Lessons from production 18:27 Results: under five minutes and 70% faster 19:48 The agent investigates and the engineer decides 20:31 Closing Sujith Nyarakkad Sudhakaran is a Senior Software Engineer at eBay who designs AI agents for monitoring and incident triage. Leonel Candido da Silva is a Senior Engineering Manager at eBay who leads teams building platforms across data engineering, machine learning, knowledge graphs, and AI agents for business-critical systems. Agent Conf brought engineers and builders together in Warsaw, Poland, on September 17 and 18, 2026 to explore the agentic way of building software. See you at another Agent Conf 🎟️ https://clstk.com/4rrdiXJ Follow Agent Conf on X to stay up to date 🐦 https://x.com/AgentConf Check out Callstack content 📚 https://clstk.com/4hgag3F #AgentConf #AgenticDevelopment #AIAgents