One thing that's becoming clear with AI agents is that traditional infrastructure monitoring isn't enough anymore. Knowing that a service is up doesn't tell you whether an agent is stuck in a retry loop, burning tokens on low-value tasks, repeatedly failing tool calls, or taking actions that create operational risk.
What stood out to me here is the focus on session-level visibility and security auditing. As agents become more autonomous, observability needs to answer not just "Is the system healthy?" but also "What decisions is the system making and why?"
The teams getting the most value from agentic systems seem to be treating cost, behavior, and safety as first-class metrics alongside latency and uptime. Token spend, tool usage patterns, failed actions, and high-risk operations are increasingly becoming part of the same operational dashboard.
Agent observability feels like it's evolving into its own discipline rather than just an extension of traditional application monitoring.