AGENTIC LARGE LANGUAGE MODELS: A SYSTEMATIC REVIEW OF ARCHITECTURES, REASONING, AND MULTI-AGENT COLLABORATION

Le résumé fourni par la source

Agentic large language models (LLMs)-systems that couple a foundation model with an execution loop for perception, planning, memory, tool use, and multi-step action-have moved from research prototypes to widely deployed infrastructure since 2022.This review synthesizes the 2023-2026 literature on agentic LLMs across four dimensions: architectures (single-and multi-agent, monolithic and modular), reasoning and planning (chain-of-thought, ReAct, Tree-of-Thoughts, and reflection-based self-correction), tool use (function calling, retrieval, and post-execution verification), and multi-agent collaboration (centralized, decentralized, and hierarchical coordination).Using a review protocol modeled on PRISMA guidance, we screened literature from arXiv, ACL/ICLR/NeurIPS proceedings, and primary industry technical reports, applying explicit inclusion and exclusion criteria and developing a structured taxonomy through iterative coding of fifty primary sources.We synthesize empirical evidence on when multi-agent systems outperform single agents-and when they do not: large-scale failure analyses find that coordination and verification failures, rather than base-model capability alone, account for the majority of multi-agent task failures.We map the benchmark landscape (GAIA, SWEbench, WebArena, AgentBench) and emerging "time-horizon" metrics that track rapid, measurable growth in the length of tasks agents can complete autonomously.We then examine open challenges in reliability, safety, and governance before turning to the technical and policy levers-compute, full-stack export strategy, standards leadership, and talent-most relevant to sustaining U.S. leadership in agentic artificial intelligence amid intensifying international competition.We close with a forward-looking research agenda spanning verifiable evaluation, self-improving systems, and multi-agent security.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.