Agentic AI Frameworks: Orchestration Architectures and Autonomous LLM-Based Reasoning Systems
Фреймворки агентного ИИ: архитектуры оркестрации и автономные системы рассуждения на основе больших языковых моделей
2026-06-11
SCID: 54.1/yvzcqjwy
Discuss with AI
Agentic AI systemsHallucination mitigationLLM-based reasoningMulti-agent architecturesReAct reasoning
Figures from the paper
Abstract (AI)
Agentic AI systems shift from static generative models to autonomous agents capable of reasoning and action. However, research remains fragmented across definitions, architectures, and evaluation practices, limiting informed adoption. Methods: 20 publications were reviewed (2021-2025) analyzing reasoning paradigms (CoT, ReAct, ToT), orchestration models (single-agent to multi-agent), memory designs, and evaluation methodologies using a five-dimensional analytical framework. Results: ReAct-based loops dominate (68 % adoption). Only 32% report safety mechanisms; evaluation is largely qualitative. Multi-agent architectures improve robustness but introduce 2-5x latency; hierarchical hallucination mitigation reduces errors by$\mathbf{4 0 - 5 0 \%}$. Conclusion: The proposed research provide practitioners a structured framework for selecting reasoning, orchestration, and guardrail strategies. Key gaps remain in standardized benchmarks, safety governance, and interoperability, outlining opportunities for production-grade agentic systems.tonomous Reasoning, Multi-Agent Systems, Tool-Augmented Intelligence
Key Findings
1
A review of 20 publications from 2021–2025 applies a five-dimensional framework to analyze agentic AI reasoning, orchestration, memory, and evaluation.
2
Hierarchical hallucination-mitigation strategies reduce errors by 40–50%, while standardized benchmarks, safety governance, and interoperability remain major gaps.
3
Multi-agent architectures improve robustness but incur a 2–5× latency increase compared with simpler configurations.
4
Only 32% of studies report safety mechanisms, while evaluation practices remain predominantly qualitative.
5
ReAct-based reasoning loops are the dominant paradigm, appearing in 68% of the reviewed publications.
Research Object
agentic AI systems, including autonomous LLM-based agents and single- or multi-agent orchestration architectures
Research Subject
reasoning paradigms, orchestration models, memory designs, safety mechanisms, and evaluation practices governing the autonomy, robustness, latency, hallucination mitigation, and adoption of agentic AI systems
Publication Details
Publication Date
2026-06-11
Journal
Publisher
ISSN
Cited by
1
Access Type
Author Information
Download PDF
Subscribe to digest