Current AI assistants remain fundamentally reactive—they respond to explicit user commands but lack the capacity for autonomous goal monitoring, decomposition, and persistent follow-through. This limitation becomes especially apparent when assisting users with executive dysfunction, where the core challenge is not knowing what to do but initiating and maintaining goal-directed behavior over time. We present pf-cortex, a dual-loop control architecture inspired by prefrontal cortex function that enables AI agents to proactively manage hierarchical goal structures: (1) a Clarify Loop performing impasse-gated goal decomposition, and (2) a Commit Loop maintaining value-weighted persistence until completion. The design incorporates precision-weighting from active inference, impasse-triggering from cognitive architectures (ACT-R/Soar), and explicit safeguards against burnout and avoidance entrenchment. We validate the theoretical foundations against five established frameworks (Carver-Scheier cybernetic control, Friston's free energy principle, Koechlin's PFC hierarchy, Zeigarnik's memory accessibility research, and production-system architectures) and propose a three-layer validity model. The specification includes concrete termination criteria, NetGain formulas, and cost-efficiency strategies via hybrid rule-ML-LLM pipelines. This work bridges cognitive science theory and practical AI agent design.
Talha Orak (Fri,) studied this question.