Problem

The field of Artificial Intelligence is rapidly transitioning from generative models that produce static outputs to "Agentic AI," where systems operate as autonomous entities capable of perceiving, reasoning, planning, and executing actions within dynamic environments. This shift leverages Large Language Models (LLMs) as cognitive controllers, integrating memory, tool use, and environmental feedback to pursue complex goals. However, the rapidly evolving landscape, with designs ranging from simple single-loop agents to intricate hierarchical multi-agent systems, presents significant challenges in terms of understanding and evaluating these diverse architectures. Furthermore, deploying agentic AI introduces new risks such as "hallucination in action" (where incorrect LLM outputs lead to real-world failures), infinite loops, and indirect prompt injection, necessitating robust security and reliability measures.

Method

This paper addresses the challenges by proposing a unified, architecture-focused taxonomy for LLM-based agents, breaking them down into six modular dimensions: Core Components (perception, memory, action, profiling), Cognitive Architecture (planning, reflection), Learning, Multi-Agent Systems, Environments, and Evaluation. This systematic approach allows for an in-depth analysis of how concrete agent systems are built, deployed, and assessed. The method emphasizes an engineering and systems perspective, examining practical design choices such as memory backends, agent computer interfaces (like "code as action" and "computer use"), standardized connector layers (e.g., Model Context Protocol - MCP), and orchestration controllers that enforce typed state and explicit transitions. For multi-agent systems, the paper analyzes interaction patterns like chain, star, and mesh topologies, and frameworks such as CAMEL, AutoGen, and MetaGPT, moving towards controllable workflow graphs rather than unstructured chat loops. The evaluation methodology adopts the "CLASSic" framework (Cost, Latency, Accuracy, Security, Stability) to assess agent performance and safety in realistic deployment scenarios.

Results

The investigation reveals a significant evolution in agentic architectures. Perception modules have advanced from text-only inputs to multimodal (visual, audio, 3D geometry), enabling agents to interpret complex user interfaces. Memory systems now support persistent state through vector databases and hierarchical summarization, overcoming context window limitations. Action spaces have expanded from fixed API calls to flexible "code as action" and direct "computer use" (mouse/keyboard operations). Cognitive architectures have progressed from linear planning (ReAct) to hierarchical and search-based methods (Tree of Thoughts, ReAcTree), complemented by reflection mechanisms for self-correction. Learning has evolved from in-context prompting to permanent weight updates through agent tuning and scalable oversight with AI feedback (RLAIF). Multi-agent systems demonstrate enhanced collaboration through role-playing, dynamic agent networks, and structured Standard Operating Procedures (SOPs), often organized into explicit workflow graphs. Agents are successfully operating across diverse environments, including web browsers, operating systems, embodied robotics (e.g., Minecraft, real-world manipulation, autonomous driving), and specialized domains like scientific discovery, healthcare, and finance. Evaluation practices now consider a holistic "CLASSic" framework, highlighting trade-offs between cost, latency, accuracy, security (especially against prompt injection), and stability.

Implications

The advancements in Agentic AI promise to transform AI systems into active collaborators capable of automating complex workflows across various domains. However, significant open challenges remain. "Hallucination in action" and persistent "infinite loops" in agent behavior necessitate more robust error recovery and meta-cognitive modules. The computational overhead of deep reasoning and multi-agent coordination requires efficient "System 2" thinking to be distilled into "System 1" reflexes for real-time applications. Ensuring social alignment and adherence to human values is crucial as agents gain more autonomy, requiring constitutional AI frameworks beyond simple RLHF. Future research directions include developing agents capable of open-ended self-improvement, continuously acquiring and refining skills without constant human intervention. Ultimately, progress will depend on integrated architectures that prioritize not only power but also controllability, auditability, and alignment with the constraints of real-world deployment, moving beyond mere model scale to complete, robust agent systems.