Synthetic Agency Architecture (SAA): Unifying Intentionality, Meta-Cognition, and Resource-Aware Adaptation in Artificial Cognitive Systems

This research on Synthetic Agency Architecture (SAA) establishes a comprehensive theoretical and computational foundation for creating artificial cognitive systems capable of purposive, self-regulating behaviour that is both transparent and contextually adaptive. Unlike fragmented approaches that address perception, planning, or learning in isolation, SAA introduces a unified hierarchical framework enabling a dynamic balance between structured goal-directed reasoning and the flexibility required to navigate complex, unpredictable environments. Grounded in intentionality theory, meta-cognitive science, and resource-bounded rationality, SAA integrates: An Intentional Layer (IL): An explicit representational system encoding agent goals, plans, and commitments as formal belief-desire-intention structures anchored in deontic and propositional logic. This ensures that decisions are traceable to a clear, auditable goal foundation, with a learned plan evaluation function replacing brittle hand-crafted plan libraries. A Meta-Cognitive Layer (MCL): A monitoring and control mechanism that continuously evaluates performance, epistemic confidence, and cognitive load across all lower processing layers. Drawing on dual-process theory, the MCL dynamically switches between fast approximate and slow deliberate reasoning strategies, triggering goal revision, memory consolidation, and plan reconsideration as environmental conditions demand. A Resource Allocation Layer (RAL): A learning component that dynamically distributes computational effort across architectural modules based on task urgency and environmental volatility. This adaptive mechanism grants the system the ability to remain efficient under resource constraints while concentrating cognitive investment precisely where marginal returns are highest. Empirical simulations across four benchmark domains — sequential decision-making, open-ended natural language interaction, multi-agent coordination, and embodied navigation — demonstrate that SAA-enabled agents achieve up to 34% higher task success rates and 33% greater resource utilisation efficiency compared to both classical BDI agents and state-of-the-art deep reinforcement learning baselines. Furthermore, SAA's adaptive resilience mechanism allows it to recover from mid-episode environmental perturbations more than twice as fast as competing architectures, effectively bridging the gap between rigid symbolic planning and flexible learned behaviour. SAA thus pioneers a new synthetic science of cognitive agency — one that moves beyond narrow task-optimised or opaque black-box models toward artificial systems capable not only of goal-directed reasoning, but also of reflecting on their own processes, justifying their decisions, and adapting to the shifting demands of the real world.