Agentic systems can fail, slow down, or overspend in ways that are difficult to diagnose from the final response alone. A single request may involve planning, retrieval, model calls, tool execution, agent handoffs, retries, policy decisions, and evaluation steps. Without end-to-end telemetry, teams cannot reliably determine where quality was lost, latency accumulated, or tokens were wasted.
This session demonstrates how OpenTelemetry can provide a common observability model across the agent lifecycle. It shows how traces and spans can represent prompts, model selections, retrieved context, tool calls, agent transitions, retries, token consumption, estimated cost, latency, evaluation results, and final business outcomes. These signals can expose unnecessary reasoning loops, oversized context, redundant tool calls, ineffective retrieval, expensive model choices, and recurring failure paths.
The session then connects production telemetry to the development lifecycle. Trace-derived findings can generate optimization proposals, create regression tests, inform prompt and workflow revisions, and pass through evaluation, artifact, approval, canary, and rollback gates before release. This creates a governed improvement loop rather than allowing agents to modify production workflows without oversight.
Attendees will learn how to use telemetry to improve reliability, adapt routing and context strategies, reduce autonomous-agent costs, and establish measurable evidence that each workflow revision is safer, faster, or more effective than the version it replaces.