Behavioral Health Control Engineering (Rivera/Hekler/Murphy, 2007-2026)
Status: NOT a competitor. Different machine, different control topology, different objective function.
What they do
Across ~20 papers examined (Hekler et al. 2018, Rivera et al. 2007, Nahum-Shani et al. 2017, van Genugten et al. 2025, Lutzow et al. 2025): every system puts the machine in controller role, optimizes for behavioral targets (steps/day, cigarettes/day) or symptom scores (PHQ-9, GAD-7). Capacity-adjacent variables (self-efficacy, stress) appear as mediators or disturbance inputs, never as terminal optimization target.
The machine: detects signals -> applies decision rules -> delivers pre-programmed interventions. It does not understand what the person values. It does not recognize narrative divergence. It cannot modulate its own communication based on deep context.
What transfers
Control theory application to human behavior, system identification methods, wearable sensing, RL-based optimization.
What does NOT transfer
Topology, objective function, machine intelligence requirements.
Positioning statement
An established control engineering program for adaptive behavioral interventions exists. It controls behavior. AgentSee proposes to control the precondition for behavior: the human's capacity to direct their own action. The engineering methods transfer. The architecture, the objective function, and the requirements on the machine are fundamentally different -- because the machine specified here requires AI understanding capabilities that are new.
Sources
- Hekler et al. 2018 (J Med Internet Res 20(6):e214)
- Nahum-Shani et al. 2017 (Ann Behav Med 52:446-462)