A comprehensive systematic literature review examining the rapid emergence of agentic workflows in Large Language Models reveals that collaborative multi-agent organizations have become the most prevalent empirical architecture in academic research, even as true operational autonomy remains constrained by pipeline-level vulnerabilities. Published in Discover Computing by Vivek Bhardwaj, Deepak Thakur, Balwinder Kaur Dhaliwal, Jeyaganesh Viswanathan, and Nandini Babbar, the investigation synthesized 126 Scopus-indexed, peer-reviewed journal articles published between 2023 and June 2026. The authors observed an extraordinary concentration of scholarly output in recent months, with 122 of the 126 analyzed studies appearing across 2025 and the first half of 2026 alone. Rather than characterizing this acceleration as unprecedented without comparative cross-discipline bibliometrics, the researchers documented the descriptive surge from a single publication in 2023 and three in 2024 to 39 in 2025 and 83 during the first six months of 2026. Their findings demonstrate that while language models have successfully transitioned from passive, single-turn text generators into dynamic systems that plan tasks, retain memory, invoke external tools, and coordinate with peer agents, the vast majority of current implementations remain tightly bounded research prototypes rather than dependable production deployments.
To evaluate this rapidly developing engineering landscape with rigorous methodological transparency, the authors structured their systematic investigation under the PRISMA 2020 guidelines around seven fundamental research questions. An initial comprehensive query across the Scopus database targeting agentic systems, foundation models, and workflow orchestration retrieved 2,309 bibliographic records, which were subsequently refined by publication stage, source type, and language into 620 peer-reviewed English-language journal articles. Following title and abstract screening, 240 papers advanced to full-text assessment, yielding a final corpus of 126 studies that met all inclusion criteria. Each included paper underwent a formal quality assessment across seven distinct technical criteria, classifying 82 studies as high quality with scores of 5.5 or greater out of 7, 38 studies as moderate quality, and six as acceptable quality. The researchers explicitly acknowledged that conducting title, abstract, and full-text eligibility screening using a single reviewer introduces an inherent risk of selection bias and subjective misclassification, while restricting the corpus strictly to peer-reviewed journal articles excluded nascent algorithmic breakthroughs, frameworks, and bleeding-edge benchmarks that frequently debut in conference proceedings or preprints.
Addressing agentic architecture, the review introduced a multi-layered taxonomy that strictly decouples primary coordination topologies from functional specialization, cross-cutting capability layers, and governance mechanisms. Prior surveys often relied on flat, mutually exclusive categories that conflated communication patterns with operational roles or tooling interfaces. Across the 57 empirical studies that provided sufficient technical detail for architectural coding, 46 exhibited an identifiable dominant coordination structure, while 11 represented hybrid configurations that combined multiple mechanisms without a single dominant pattern. Collaborative multi-agent networks emerged as the most frequently observed primary topology, appearing in 27 empirical studies, which represents 47.4 percent of the empirical subset and 21.4 percent of the entire 126-paper corpus. Hierarchical supervisory frameworks, where a higher-level controller directs, evaluates, or gates subordinate execution, accounted for 10 empirical studies or 17.5 percent. Single-agent systems, relying on one central language model without autonomous delegation to secondary agents, comprised nine empirical studies or 15.8 percent. Role specialization, tool augmentation, and human-in-the-loop oversight were classified as orthogonal functional, capability, and governance attributes rather than standalone topologies.
The review systematically mapped planning and reasoning strategies into five mechanistically distinct categories defined by the structural technique used to extend, correct, or redirect an agent’s execution trajectory. Linear reasoning methods such as Chain-of-Thought decompose complex objectives into sequential steps but lack native mechanisms to explore branching alternatives or recover from early miscalculations. Search-based paradigms, including Tree-of-Thought and Graph-of-Thought frameworks such as ESCARGOT, permit agents to explore candidate paths, reuse intermediate reasoning states, and navigate interconnected dependencies, albeit at substantially higher computational and token costs. Self-improvement mechanisms, exemplified by Reflexion and self-refinement routines, enable iterative self-correction by evaluating prior actions against environmental feedback, though additional refinement cycles directly inflate execution latency. Collaborative reasoning distributes decision-making across peer agents through multi-agent debate and consensus protocols, which improve decision robustness and reduce individual bias at the expense of communication overhead. Finally, adaptive planning paradigms incorporate deliberative planning, dynamic backtracking, and runtime modifications, allowing industrial and robotic workflows to recover from execution failures and adapt to fluid operational constraints.
Investigating contextual awareness and long-horizon execution, the authors synthesized memory into four complementary functional tiers that operate as a continuous information lifecycle. Short-term working memory manages active conversational context, immediate observations, and action-observation pairs within the model context window, but remains capacity-constrained and susceptible to attention dilution. Episodic memory preserves temporally ordered traces of historical actions, errors, and outcomes, enabling systems like AutoSecAgent to learn from prior trial-and-error attempts, though long-running agents face trajectory bloat that necessitates aggressive pruning and experience ranking. Semantic memory stores reusable domain knowledge and organizational policies within persistent vector databases, facilitating targeted retrieval during execution. To overcome the relational limitations of dense vector similarity in multi-hop queries, hybrid retrieval-augmented generation architectures integrate knowledge graphs with vector embeddings, as demonstrated in the PersonalAI and ESCARGOT platforms. In multi-agent configurations, shared blackboards provide collaborative spaces for state sharing, hypothesis generation, result propagation, and conflict resolution, while demanding strict access control, versioning, and synchronization policies to prevent data corruption.
Despite these architectural advances, the review emphasized that persistent memory systems introduce substantial operational vulnerabilities and scaling degradation over extended operating horizons. Context accumulation causes expanding historical records to compete directly for limited working memory capacity, elevating inference expenses and degrading contextual attention. Similarly, retrieval noise emerges when large episodic stores return historical experiences that are semantically similar but operationally inappropriate, potentially causing agents to repeat prior failed trajectories. Environmental shifts generate knowledge drift, rendering previously stored facts, operational rules, or technical configurations obsolete. At the system level, memory poisoning poses an acute security hazard, occurring when malicious or incorrectly generated entries are inserted into persistent memory or external knowledge bases and subsequently retrieved into active reasoning contexts. Because retrieved information directly influences downstream planning and action execution, poisoned memory can generate persistent error loops across multiple future tasks. In distributed multi-agent systems, lack of consistent state updates causes memory desynchronization, leading peer agents to reason from contradictory constraints, issue conflicting tool invocations, or experience coordination deadlock.
Examining environmental interaction, the study organized tool utilization across four complementary integration layers: structured data contracts, programmatic execution sandboxes, informational interlocks, and protocol abstractions. Structured data contracts utilize typed function specifications and schema-validated JSON interfaces to translate probabilistic language outputs into deterministic API requests, bridging the operational impedance mismatch between models and external services. Programmatic sandboxes provide isolated computational execution environments where agents generate code to perform exact numerical calculations, data manipulation, or physical simulations, as seen in molecular simulation tools like Simulon and ToPolyAgent. Informational interlocks ground agent reasoning in dynamic external databases, knowledge graphs, and live search engines through SQL, Cypher, or vector queries. To mitigate the complexity of engineering bespoke integrations for expanding tool ecosystems, protocol abstractions like the Model Context Protocol standardize tool discovery, resource access, and prompt templating across client-server boundaries. However, the authors noted that tool integration magnifies security risks, requiring strict containment against prompt injection, unauthorized system access, data exfiltration, and execution-level resource exhaustion through least-privilege scoping and isolated sandboxing.
Across real-world application domains, the review identified significant disparities in experimental maturity, driven primarily by the environmental embodiment and operational consequences of agent actions. Digital and information-intensive environments currently demonstrate the highest degree of experimental readiness, led by software engineering. In coding tasks, the availability of compilers, test suites, and version-control systems provides immediate, objective, and executable feedback, allowing agents to iteratively diagnose and repair bugs while offering low-cost action reversibility. Cybersecurity workflows exhibit substantial promise for automated threat intelligence, log parsing, and controlled penetration testing, but remain constrained by adaptive human adversaries and high operational risks. Scientific discovery leverages computational sandboxes to accelerate hypothesis generation and molecular modeling, yet remains bounded by the time and financial expense of downstream physical laboratory verification. Educational tutoring workflows benefit from longitudinal learner memory and interactive sandboxes, but their deployment remains restricted by pedagogical validity, privacy considerations, and the risk of hallucinated feedback reinforcing student misconceptions.
Conversely, high-consequence decision environments and embodied cyber-physical systems exhibit much lower readiness for unrestricted autonomous operation. In healthcare and clinical decision support, where erroneous recommendations directly endanger patient safety, the literature universally mandates human-in-the-loop governance, multimodal evidence grounding, and rigorous diagnostic calibration rather than autonomous execution. Financial workflows similarly demand deterministic risk boundaries, transaction thresholds, and transparent audit trails to prevent automated decisions from causing severe economic losses during volatile market shifts. In robotics and industrial automation, language model inference is neither sufficiently deterministic nor fast enough to execute real-time physical control safely. Consequently, embodied architectures universally decouple high-level semantic task planning handled by the language model from low-level deterministic controllers, such as Model Predictive Control, which enforce physical kinematic bounds and collision avoidance. Industrial manufacturing deployments further encounter severe infrastructure bottlenecks, where semantic agents must interface with heterogeneous legacy machinery, proprietary communication protocols, and stringent availability standards, positioning digital twins as essential intermediate validation environments.
Synthesizing evaluation and benchmarking methodologies, the authors argued that assessing multi-step agentic workflows requires evaluating entire sequential execution trajectories rather than measuring terminal task completion in isolation. Two workflows achieving identical terminal success may display vastly divergent step efficiency, tool grounding accuracy, and recovery behaviors. Modern evaluation protocols therefore employ multi-dimensional trajectory metrics, including the Plan Optimality Ratio, which calculates the mathematical ratio between the minimal viable execution path and the actual executed trajectory to penalize state oscillations and redundant loops. Benchmarking architectures increasingly adopt a two-layer structure, combining a domain-agnostic core that measures planning efficiency, tool selection precision, memory stability, and token expenditure with domain-specific criteria such as diagnostic calibration or kinematic safety bounds. To progress beyond brittle research prototypes, the review outlined six critical research pillars, prioritizing standardized long-horizon benchmarking, capability-based security architectures, privacy-preserving ephemeral memory, auditable execution logging, unified multi-agent governance, and bounded self-evolution to establish verifiable safety before expanding operational autonomy.
Subject of Research: Systematic literature review of architectures, reasoning strategies, memory structures, and deployment limitations in large language model agentic workflows
Article Title: Agentic workflows in large language models: a comprehensive review of architectures, planning strategies, and real-world applications
Article References: Bhardwaj, V., Thakur, D., Dhaliwal, B. K., Viswanathan, J., & Babbar, N. (2026). Agentic workflows in large language models: a comprehensive review of architectures, planning strategies, and real-world applications. Discover Computing, 29(1), Article 908. https://doi.org/10.1007/s10791-026-10743-4
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10743-4
Keywords: agentic workflows, large language models, multi-agent systems, systematic literature review, PRISMA 2020, hierarchical supervisory architectures, Chain-of-Thought, Tree-of-Thought, episodic memory, hybrid RAG, Model Context Protocol, tool augmentation, memory poisoning, software engineering agents, clinical decision support, trajectory evaluation, Plan Optimality Ratio, human-in-the-loop governance
News Source: Denise Maddox. (November 18, 2026). Systematic Review Finds Collaborative Multi-Agent Networks Dominate Research While Autonomy Remains Constrained. Scienmag.



