Who's monitoring the agents?

Multi-agent AI systems built with frameworks like CrewAI, AutoGen, and LangGraph are moving into production, but operational visibility remains severely lacking. Teams are deploying these systems with less monitoring than they had for microservices a decade ago. The core problems include runaway model call chains that inflate latency and cost without triggering alerts, subtle failures buried in agent decision chains, and gradual data propagation across agents that can cross security boundaries unnoticed. Traditional logs and traces don't capture how an agent system actually arrived at an outcome. What's needed is visibility into the full execution graph — how requests unfold across agents, where reasoning chains branch or loop, and how data transforms across steps. The author argues that agent systems develop behavioral baselines over time, and effective monitoring should detect deviations from normal patterns rather than relying on static rules.