Work

Agentic AI: A Review and Proposed Framework

Year: 2026
Type: article
Abstract: Chatbots built on large language models (LLMs) are good at producing text, but most of them stop at the reply. They cannot look up an order, run a test or update a record unless someone connects them to tools and tells them how to use those tools safely. Agentic AI is the name given to systems that close this gap: an LLM is placed inside a loop where it makes a plan, calls a tool, reads what came back and decides what to do next. This paper reviews the main building blocks of such systems, namely planning and reasoning, tool use, memory and state, and human oversight, and places them in the recent literature on ReAct-style agents, tool learning, reflection, multi-agent conversation and agent benchmarks. Building on this review, we propose a modular agent pipeline made of a tool registry with permissions, a planner, a permission check, an optional human approval gate, tool execution with observation, result checking and an audit log. We also describe how such a pipeline should be evaluated, using task success, tool errors and injection resistance, and time and cost. The paper is a review with a proposed design and does not report new experiments. It closes with open problems, including reliable planning over long tasks, defence against prompt injection, and fair evaluation of agents that behave differently on every run.
Language: English
Open Access status: diamond
OpenAlex ID: W7226005145
FWCI:
Cites:
Cited by:
Related to: 10