Comprehensive review uncovers fundamental safety, privacy, and design challenges in agentic artificial intelligence, highlighting essential pathways toward reliable autonomous systems.
Key Points
To synthesize foundational concepts, propose a structural taxonomy, and outline safety and evaluation frameworks for developing trustworthy agentic artificial intelligence.
Synthesized insights from philosophy, cognitive science, and machine learning to characterize agency across reinforcement learning, symbolic reasoning, and embodied cognition paradigms.
Developed a multi-dimensional taxonomy classifying systems across autonomy, cognitive capability, modality, and environmental interaction.
Conducted a systematic analysis of enabling technologies, benchmark suites, verified autonomy methods, and alignment strategies.
Identified critical safety, alignment, and privacy vulnerabilities across autonomous agent workflows, demonstrating the necessity of functional closed-loop behavioral evaluations over static testing.