Agentic Presence Architecture and Deployment Audit for Proactive Embodied Agents in Extended Reality

Proactive embodied agents in extended reality (XR) require an explicit interface between inferred user need and embodied intervention. We present Agentic Presence, a six-layer reference architecture integrating multimodal context, intent inference, local initiative arbitration, anticipatory scheduling, and embodied actuation. Its audit separates theoretical decision regions, sensor reachability, and decisions issued under runtime constraints. A Unity implementation on tethered Meta Quest 3 ran the arbiter at a configured 10 Hz, with median execution of four microseconds and decision-to-cue latency of at most 1 ms. Across 19,929 agentic-condition ticks, the audit identified cadence and speech gates as the dominant constraints. The system issued 75 interventions, comprising 44 ACT, 31 ASK, and no OFFER actions; the recovering presence-risk proxy remained at one on 98.5% of logged ticks. This separation established which mechanisms the deployment exercised and which required different task conditions. A within-subject pilot (n = 13) estimated the complete initiative package relative to an identical reactive companion: a co-presence difference of 0.64 points (95% CI [−0.11, 1.40]; dz = 0.51), with a one-sided 95% upper bound of 0.70 points for increased intrusiveness. These preliminary estimates do not isolate arbitration timing or validate the risk proxy. The contribution is an implementable architecture and a reproducible deployment audit that connects formal policy design to realized sensing and runtime behavior.