Agentic workflows in large language models: a comprehensive review of architectures, planning strategies, and real-world applications

Large Language Models (LLMs) are increasingly deployed as agentic workflows: systems in which an LLM plans tasks, retains memory, uses external tools, and coordinates with other agents to complete multi-step objectives. Existing surveys typically evaluate these components in isolation or rely on flat, mutually exclusive categorizations that conflate coordination patterns with functional roles and tooling. Following PRISMA 2020 guidelines, this systematic review synthesizes 126 Scopus-indexed, peer-reviewed journal articles published between 2023 and June 2026, with screening conducted by a single reviewer. Because 122 of these articles appeared in 2025 and 2026, the corpus provides a timely, peer-evaluated evidence base for this rapidly accelerating field. Across seven research questions, we introduce a multi-layered taxonomy whose explicit novelty lies in strictly decoupling three primary coordination topologies (single-agent, collaborative multi-agent, and hierarchical supervisory) from orthogonal functional dimensions (role specialization), cross-cutting capability layers (tool augmentation and hybrid memory), and governance mechanisms (human-in-the-loop). Unlike prior rigid taxonomies that force systems into a single category, this structural separation systematically models modern hybrid implementations by combining distinct operational layers without artificial pigeonholing. Among the 57 empirical studies, collaborative multi-agent organization was the most frequently observed identifiable primary topology (27 studies, representing 47.4% of the empirical subset [ n = 27/57] and 21.4% of the total 126-study corpus). Across software engineering, healthcare, cybersecurity, and robotics, most reported implementations remain controlled research prototypes. Practical autonomy continues to be constrained by pipeline-level vulnerabilities: compounding error propagation, indirect prompt injection, memory poisoning, and the lack of standardized trajectory-level benchmarks. To advance agentic workflows toward dependable deployment, future research must prioritize standardized long-horizon evaluation, capability-based security architectures, and controlled self-evolution.

Authors

Institutions

Publication Details

Journal
Discover Computing
Published
2026-11-18
DOI
https://doi.org/10.1007/s10791-026-10743-4
Primary Topic
AI-based Problem Solving and Planning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Agentic workflows in large language models: a comprehensive review of architectures, planning strategies, and real-world applications

Deepak S. Thakur, Vivek Bhardwaj, Balwinder Kaur Dhaliwal, Nandini Babbar et al.
1 links
Discover Computing
AI-based Problem Solving and Planning
article

Agentic workflows in large language models: a comprehensive review of architectures, planning strategies, and real-world applications

Deepak S. Thakur, Vivek Bhardwaj, Balwinder Kaur Dhaliwal, Nandini Babbar, Jeyaganesh Viswanathan
article en

Abstract

Large Language Models (LLMs) are increasingly deployed as agentic workflows: systems in which an LLM plans tasks, retains memory, uses external tools, and coordinates with other agents to complete multi-step objectives. Existing surveys typically evaluate these components in isolation or rely on flat, mutually exclusive categorizations that conflate coordination patterns with functional roles and tooling. Following PRISMA 2020 guidelines, this systematic review synthesizes 126 Scopus-indexed, peer-reviewed journal articles published between 2023 and June 2026, with screening conducted by a single reviewer. Because 122 of these articles appeared in 2025 and 2026, the corpus provides a timely, peer-evaluated evidence base for this rapidly accelerating field. Across seven research questions, we introduce a multi-layered taxonomy whose explicit novelty lies in strictly decoupling three primary coordination topologies (single-agent, collaborative multi-agent, and hierarchical supervisory) from orthogonal functional dimensions (role specialization), cross-cutting capability layers (tool augmentation and hybrid memory), and governance mechanisms (human-in-the-loop). Unlike prior rigid taxonomies that force systems into a single category, this structural separation systematically models modern hybrid implementations by combining distinct operational layers without artificial pigeonholing. Among the 57 empirical studies, collaborative multi-agent organization was the most frequently observed identifiable primary topology (27 studies, representing 47.4% of the empirical subset [ n = 27/57] and 21.4% of the total 126-study corpus). Across software engineering, healthcare, cybersecurity, and robotics, most reported implementations remain controlled research prototypes. Practical autonomy continues to be constrained by pipeline-level vulnerabilities: compounding error propagation, indirect prompt injection, memory poisoning, and the lack of standardized trajectory-level benchmarks. To advance agentic workflows toward dependable deployment, future research must prioritize standardized long-horizon evaluation, capability-based security architectures, and controlled self-evolution.

Discover ComputingVol. 29(1)
Lovely Professional University (IN), Amity University (IN), Zoetis (United States) (US), Manipal University Jaipur (IN)
AI-based Problem Solving and Planning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon