Metraipolis › archive › dc_7946c5fca7ae

Agentic AI: From Large Language Models to Autonomous Tool-Using Agents (2026) · OpenAlex W7168281372

Source
OpenAlex
Original
https://doi.org/10.5281/zenodo.21347786
License (as archived)
CC-BY-4.0
Archived
2026-10-07T09:02:37.710Z
SHA-256 of the archived text
e2b43857686e779e5bee3777c0598352e1614c9fc1883c59b00681ba38f76816
Stable URL of this page
https://metraipolis.com/documento/dc_7946c5fca7ae

How to cite

Agentic AI: From Large Language Models to Autonomous Tool-Using Agents (2026) · OpenAlex W7168281372 — OpenAlex, https://doi.org/10.5281/zenodo.21347786 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_7946c5fca7ae (Metraipolis archive dc_7946c5fca7ae, sha256 e2b43857686e779e5bee3777c0598352e1614c9fc1883c59b00681ba38f76816)
BibTeX
@misc{dc7946c5fca7ae,
  title = {Agentic AI: From Large Language Models to Autonomous Tool-Using Agents (2026) · OpenAlex W7168281372},
  howpublished = {OpenAlex},
  url = {https://doi.org/10.5281/zenodo.21347786},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_7946c5fca7ae; sha256 e2b43857686e779e5bee3777c0598352e1614c9fc1883c59b00681ba38f76816},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: OpenAlex, https://doi.org/10.5281/zenodo.21347786. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Opera scientifica in accesso aperto: scheda dai metadati OpenAlex (other).
Titolo: Agentic AI: From Large Language Models to Autonomous Tool-Using Agents
Autori: Abeda Elhenawy
Anno: 2026 · Sede: Zenodo (CERN European Organization for Nuclear Research)
DOI: https://doi.org/10.5281/zenodo.21347786
Accesso aperto: green · testo: https://doi.org/10.5281/zenodo.21347786 · licenza del testo: cc-by
Citazioni ricevute (OpenAlex): 0
Temi (OpenAlex): AI-based Problem Solving and Planning; Artificial Intelligence Applications; Natural Language Processing Techniques

Abstract (testo degli autori, licenza cc-by):
Abstract The emergence of large language models (LLMs) has fundamentally reshaped the landscape of artificial intelligence, but the transition from passive text generation to active, goal-directed intervention marks a more profound paradigm shift. This paper presents a comprehensive survey of agentic AI systems that integrate LLMs with perception, memory, planning, and tool-use to autonomously pursue objectives across digital and physical environments. We propose a taxonomy that organizes agentic systems along four orthogonal dimensions (single- vs. multi-agent, reactive vs. deliberative, open- vs. closed-loop, and domain-general vs. domain-specific), and we dissect the core architectural components that underpin modern agent frameworks: stratified memory, planning paradigms (ReAct, Chain-of-Thought, Tree-of-Thoughts, Reflexion), tool calling, and multi-agent choreography. We synthesize the rapidly growing benchmark landscape (SWE-bench, WebArena, GAIA, AgentBench, ToolBench) and survey deployments across software engineering, scientific discovery, enterprise automation, and personal assistance. We further analyze the principal open problems—hallucination under agentic load, alignment drift, prompt-injection attack surfaces, the evaluation crisis, continual learning, and the societal renegotiation of human agency—and outline future directions centered on world models, causal reasoning, standardized evaluation, and governance frameworks. The survey covers literature from January 2022 through July 2026 and is structured to serve as a reference for researchers, practitioners, and policymakers engaging with the agentic AI frontier. Keywords: Agentic AI; Large Language Models; Autonomous Agents; Tool Use; Multi-Agent Systems; Planning and Reasoning; Alignment; Survey. 1 Introduction The rapid maturation of large language models (LLMs) following the introduction of the transformer architecture and the demonstration of emergent in-context capabilities at scale has prompted a fundamental rethinking of what artificial intelligence systems can do. Whereas the previous decade of deep learning was dominated by discriminative models trained for narrow tasks—image classification, speech recognition, sentiment analysis—the current decade is increasingly defined by generative models that produce fluent text, code, and structured reasoning across domains they were never explicitly trained on. Yet fluency alone does not amount to intelligence in any practically useful sense. A model that merely predicts the next token, however skilfully, remains a passive instrument: it generates, but it does not act. The shift from passive generation to active intervention is the defining move of agentic AI. An AI agent, in the sense now crystallising across the research literature, is a system that perceives its environment, reasons about what it has perceived, and takes actions that reshape that environment in pursuit of a goal. The distinction between an LLM and an LLM-based ...

Fonte: https://openalex.org/W7168281372 — metadati OpenAlex (CC0 1.0), consultati il 2026-10-06 via api.openalex.org. Interrogazione: «large language model agents tool use».