The defining workload of 6G is expected to be machine-native intelligence rather than human-driven data transfer, yet no study has measured the arrival process a deployed autonomous agent generates on a real radio link. This letter contributes the first empirical characterization of real agent traffic from a deployed platform over a real cellular link (extending prior controlled testbed measurements to a live wireless edge client), observed via application-layer telemetry, which captures the coupled generation–transmission process without packet capture. We propose a two-level framework—withintask (Level 1) versus cross-task pooled (Level 2)—and show that the pooled heavy tail is quantitatively explained by log-variance additivity, dominated by within-task log-variance (∼ 83%) with a smaller cross-task heterogeneity contribution (∼ 17%). Maximum-likelihood fitting with KS and AIC selects log-normal inter-arrival times and decisively rejects Poisson (). The findings yield quantitative, p99-based uplink capacity-planning rules for 6G edge clients.