Session traces and cost controls are emerging as key observability techniques for diagnosing AI agent failures, helping teams spot tool-call loops and runaway spend while preserving enough execution context for post-incident debugging. By Mark Silvester