Adapt-Plan: A Hybrid Control Architecture for PEI-Guided Reliable Adaptive Planning in Dynamic Agentic Environments

The field of agentic artificial intelligence is transitioning from reasoning-centric architectures toward systems explicitly designed for reliability under uncertainty. Current agent frameworks, while effective in controlled environments, exhibit cognitive rigidity—an inability to proactively correct planning trajectories when confronted with unexpected faults. This paper introduces Adapt-Plan, a foundational hybrid architecture that reformulates planning as a control-theoretic process by elevating the Planning Efficiency Index (PEI) from a post-hoc evaluation metric to a real-time control signal. Through dual-mode planning (strategic and tactical) and Extended Dynamic Memory (EDM) for selective experience consolidation, Adapt-Plan enables agents to detect behavioral drift early and initiate corrective adaptation before catastrophic degradation occurs. Controlled validation across 150 episodes demonstrates PEI=0.91 ± 0.06 and FRR=78% ± 4.2% (95% CI [74%, 82%], p < 0.001, Cohen’s d = 2.18 vs. ReAct), establishing the algorithmic viability of metric-driven adaptation. Comprehensive ablation studies isolate component contributions, revealing that PEI-guided control accounts for 31% of performance gains. These architectural principles were subsequently validated at scale through rigorous certification frameworks, confirming that PEI-driven control generalizes to deployment-grade reliability when augmented with safety protocols. This work establishes the conceptual foundation for reliable agentic AI through the tight integration of architecture, metrics, and control.