Agentic Monitoring in Production
summary
eBay deploys two production agents — a diagnosis agent that turns docs and runbooks into a structured investigation plan and an enrichment agent that queries live metrics, correlations and segments — to answer “should this alert be escalated?” within five minutes. They trade heavy LLM context for a two-step workflow, choose a quicker GPT-4 endpoint for latency, build traceability to source dashboards, and report a 70% faster investigation with the engineer keeping the escalation decision.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 0:23Speakers and eBay's business reliability platform
- 0:50Anomaly detection versus investigation
- 3:02Turning alert triage into an agentic workflow
- 3:33How much should engineers trust an agent?
- 4:53Investigating an 18% revenue drop
- 5:50The manual investigation workflow
- 6:38Defining the escalation decision
- 7:28Context window limits
- 8:03Diagnosis and enrichment agents
- 9:46The end-to-end workflow
- 10:50From a raw alert to an enriched notification
- 13:02eBay Ads investigation example
- 15:02Model choice and production latency
- 16:46Building traceability into the workflow
- 17:28Lessons from production
- 18:27Results: under five minutes and 70% faster
- 19:48The agent investigates and the engineer decides
- 20:31Closing
