PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic AI: A Comprehensive Survey of Architectures, Applications, and Future Directions

Mohamad Abou Ali, Fadi Dornaika

arXiv:2510.25445v1cs.AIcs.LG

TL;DR

Agentic AI research lacks a clear account of the relationship between modern neural agents and classical symbolic frameworks, a problem that motivates this survey. The paper applies a dual-paradigm taxonomy in a systematic review to distinguish their architectures and operational mechanisms. It concludes that the paradigms are fundamentally divergent, have domain-dependent strengths, and point toward hybrid architectures with paradigm-specific governance.

  • Problem

    Existing discourse conceptually retrofits classical BDI and PPAR frameworks onto LLM-based agents, obscuring their stochastic, prompt-driven mechanics and conflating incompatible paradigms.

  • Method

    The survey uses a dual-paradigm framework to classify and compare symbolic/classical and neural/generative systems, mapping their architectures, coordination mechanisms, and governance challenges.

  • Results

    The analysis finds that symbolic and neural lineages are parallel paradigms with fundamentally different operational mechanics, and that domain requirements shape their application strengths.

  • Takeaways & Limitations

    Future progress is expected from hybrid architectures integrating symbolic reasoning with neural orchestration while preserving adaptability and improving verifiability.

  • Takeaways & Limitations

    Methodological heterogeneity and fundamentally different performance measures limit direct cross-study benchmarking, especially between paradigms.

Abstract

from arXiv · show

Agentic AI represents a transformative shift in artificial intelligence, but its rapid advancement has led to a fragmented understanding, often conflating modern neural systems with outdated symbolic models -- a practice known as conceptual retrofitting. This survey cuts through this confusion by introducing a novel dual-paradigm framework that categorizes agentic systems into two distinct lineages: the Symbolic/Classical (relying on algorithmic planning and persistent state) and the Neural/Generative (leveraging stochastic generation and prompt-driven orchestration). Through a systematic PRISMA-based review of 90 studies (2018--2025), we provide a comprehensive analysis structured around this framework across three dimensions: (1) the theoretical foundations and architectural principles defining each paradigm; (2) domain-specific implementations in healthcare, finance, and robotics, demonstrating how application constraints dictate paradigm selection; and (3) paradigm-specific ethical and governance challenges, revealing divergent risks and mitigation strategies. Our analysis reveals that the choice of paradigm is strategic: symbolic systems dominate safety-critical domains (e.g., healthcare), while neural systems prevail in adaptive, data-rich environments (e.g., finance). Furthermore, we identify critical research gaps, including a significant deficit in governance models for symbolic systems and a pressing need for hybrid neuro-symbolic architectures. The findings culminate in a strategic roadmap arguing that the future of Agentic AI lies not in the dominance of one paradigm, but in their intentional integration to create systems that are both adaptable and reliable. This work provides the essential conceptual toolkit to guide future research, development, and policy toward robust and trustworthy hybrid intelligent systems.

1 Introduction

Agentic AI shifts AI from passive, task-specific tools toward autonomous systems, but current discourse often retrofits neural architectures into mismatched symbolic frameworks. This survey introduces a dual-paradigm taxonomy that classifies systems by architecture and analyzes their mechanisms and applications on their own terms.

  • Foundations: Agentic AI systems pursue goals autonomously through capabilities including planning, contextual memory, tool use, and adaptation to environmental feedback.The paper distinguishes autonomous agents from broader Agentic AI systems that may orchestrate multiple specialized agents.
  • Foundations: Multi-agent Agentic AI systems coordinate specialized agents to solve problems too complex for a single agent.The paper illustrates this with project management, research, writing, and quality-assurance roles collaborating on one project proposal.
  • Research gap: Conceptual retrofitting misapplies classical BDI and PPAR frameworks to LLM-based agents built on stochastic generation and prompt-driven orchestration.The paper argues that this obscures LLM-agent mechanics and falsely suggests continuity between incompatible paradigms.
  • Framework: The survey’s dual-paradigm taxonomy separates Symbolic/Classical and Neural/Generative lineages, then crosses architectural paradigm with single-agent versus multi-agent coordination.The framework is intended for classification and comparison rather than depicting historical evolution.
  • Contributions: The review contributes a framework for accurate system classification, clarification of neural agency mechanisms, and paradigm-specific governance analysis.Neural frameworks achieve agency through prompt chaining and conversation orchestration rather than symbolic planning, while governance is anchored to each paradigm’s technological context.

2 Theoretical Foundations: Mapping the Dual Lineages of Agentic Intelligence

Agentic AI branches into Symbolic/Classical and Neural/Generative lineages that pursue agency through divergent mechanisms. The symbolic lineage uses explicit logic, planning, and state models, extending from fully observable MDPs to belief-based POMDPs and cognitive architectures.

  • Dual lineages: The conceptual framework treats agentic AI as two distinct architectural lineages rather than a linear progression.Both lineages pursue autonomy and goal-directed behavior, but implement them through fundamentally different mechanisms.
  • Symbolic/Classical foundations: The Symbolic/Classical lineage relies on explicit logic, algorithmic planning, and deterministic or probabilistic models.Its foundations include rule-based reasoning, explicit state representations, and engineered cognitive processes.
  • Symbolic/Classical foundations: MDPs model fully observable environments through states, actions, transition probabilities, and rewards, supporting deterministic rule-based decision-making.Their tuple is (S, A, P, R), but they lack robust reasoning under uncertainty.
  • Symbolic/Classical foundations: POMDPs add probabilistic belief states for incomplete information, enabling hidden-state inference but imposing computational overhead that limits scalability.They remain algorithmic state-estimation systems within the symbolic approach.
  • Symbolic/Classical foundations: BDI and SOAR explicitly model internal states and implement a perceive-plan-act-reflect loop, making symbolic agency powerful but brittle in complex environments.Their architecture maps human cognitive functions directly onto symbolic modules.

2.3 The Neural Lineage: Statistical Learning and Emergent Reasoning

The Neural/Generative lineage develops from statistical learning and deep reinforcement learning toward LLM-based orchestration. Its agency emerges from stochastic generation and prompt-driven coordination rather than internal symbolic logic.

  • Neural/Generative foundations: The neural lineage progresses from statistical learning toward emergent, stochastic behavior grounded in data.This trajectory culminates in the generative capabilities of large language models.
  • Deep reinforcement learning: Deep reinforcement learning learns policies directly from high-dimensional data, moving beyond hand-crafted symbolic rules.Meta-DRL extends this transition by introducing generalization across tasks through meta-optimization.
  • LLM-based agency: LLMs created the neural paradigm by providing a general-purpose substrate for statistical prediction and generative reasoning.The architectural shift is from designing cognitive agents to orchestrating generative pipelines.
  • LLM-based agency: LangChain, AutoGen, and CrewAI coordinate tasks through LLM orchestration rather than symbolic PPAR loops or BDI architectures.Pre-trained models act as central executives that coordinate tasks through different mechanisms.
  • LLM-based agency: Neural agency is described as an emergent property of prompt-driven orchestration, not a product of internal symbolic logic.This distinction marks the paper’s definitive break between the two architectural traditions.

2.4 Multi-Agent Orchestration: The Pinnacle of the Neural Paradigm

Multi-agent orchestration represents the neural paradigm’s most advanced form, combining specialized agents through dynamic routing and structured communication. Its scalability and complex problem-solving arise from system-level coordination rather than a single agent’s cognitive complexity.

  • Multi-agent orchestration: Multi-agent systems coordinate diverse, modular agents through structured communication protocols.This extends neural agency beyond isolated systems by distributing work across specialized components.
  • Learned adaptability: Meta-DRL improves adaptability across tasks through dual-loop meta-optimization, contrasting with vanilla DRL’s single-task optimization.The figure presents this as a shift from explicit programming toward learned, emergent capabilities.
  • Paradigm transition: The evolution from MDP to POMDP to LLM orchestration shows a transition from deterministic and uncertainty-aware symbolic systems to neural coordination.The mechanisms of intelligent behavior change across the sequence rather than forming a single continuous architecture.
  • Multi-agent orchestration: An LLM-based orchestrator manages context and dynamically routes specialized subtasks to other agents.The orchestrator assesses the overall goal and assigns work through a dynamic workflow.
  • Multi-agent orchestration: The architecture achieves scalability and complex problem-solving through emergent intelligence produced by a well-orchestrated system.Its capability does not depend on a single agent’s cognitive complexity.

3 Methodology

The review uses a PRISMA 2020-based, interdisciplinary process to identify and synthesize literature across symbolic/classical and neural/generative agentic AI. Its two-phase design combines systematic contemporary retrieval with separately analyzed foundational symbolic context, while acknowledging fast-changing evidence and cross-study comparability limits.

  • Review design: The methodology follows PRISMA 2020 to guide literature search, screening, evaluation, and synthesis across both agentic paradigms.The review is designed for transparency and reproducibility in a rapidly evolving field.
  • Review design: The review objectives cover paradigm classification, capability and application evolution, governance and ethics, and research gaps with future directions.These objectives structure the synthesis around both architecture and practical implications.
  • Search strategy: The search spans multiple databases and interdisciplinary fields, using separate keyword clusters for symbolic/classical and neural/generative research.ArXiv preprints were manually assessed to capture current developments in the neural paradigm.
  • Screening and inclusion: 165 records were initially identified, 120 unique records remained after deduplication, and 78 articles met inclusion criteria after screening and full-text assessment.The review included peer-reviewed publications, technical reports, and manually screened high-impact preprints meeting the stated criteria.
  • Screening and inclusion: 12 seminal symbolic papers were added separately as foundational context, producing a two-phase review rather than a purely systematic retrieval process.The supplemental works were used for historical and theoretical framing and analyzed separately from contemporary research.
  • Limitations: Proprietary neural systems, heterogeneous evaluation metrics, and hybrid architectures constrain transparency, direct cross-study benchmarking, and unambiguous paradigm classification.Some architectural details and performance metrics were incomplete or inferred from secondary sources.

4 Literature Review: A Dual-Paradigm Analysis

The review distinguishes symbolic systems, built around algorithmic reasoning and formal protocols, from neural systems, built around stochastic generation and prompt-driven orchestration. Across applications and coordination, the analysis argues that future systems should integrate both paradigms to combine adaptability with reliability.

  • Neural paradigm: Neural frameworks use dynamic context management, prompt engineering, and tool composition rather than classical cognitive loops.
  • Emerging directions: Neuro-symbolic integration aims to bridge deterministic symbolic reasoning with the adaptive, generative capabilities of neural networks.
  • Emerging directions: Decentralized agent networks use blockchain coordination to support verifiable governance, transparent decisions, and resilience against single points of failure.
  • Emerging directions: Lifelong learning addresses largely stateless LLM-based agents by adding continuous adaptation and durable knowledge retention.
  • Multi-agent coordination: Symbolic coordination provides verifiable reliability through engineered protocols, whereas neural coordination provides adaptable emergence through learned conversation patterns.
  • Application fit: Paradigm choice is dictated by domain requirements: symbolic architectures suit reliability and safety, while neural architectures suit adaptability and unstructured data.

5 Analysis of Domain-Specific Applications

Agentic AI deployments show that domain constraints shape architectural selection and implementation priorities. Symbolic, neural, and hybrid designs are used strategically across safety, compliance, adaptability, tool integration, and resilience requirements.

  • Domain applications: Healthcare favors symbolic or highly constrained deterministic systems to prioritize safety, accuracy, and auditability in regulated environments.
  • Domain applications: Education leverages neural systems to generate adaptive, personalized, and context-aware interactions that symbolic rules cannot easily pre-program.
  • Domain applications: Finance and legal applications combine neural orchestration with role-based workflows and verified retrieval to mitigate hallucination and support compliance.
  • Domain applications: Robotics pairs symbolic systems for safety-critical low-level control with neural systems for high-level coordination and adaptation.
  • Tool use and capabilities: Symbolic tool use is hard-coded and deterministic, with predefined parameters and auditable interfaces such as healthcare EHR APIs.
  • Tool use and capabilities: Neural tool use dynamically selects tools and generates API parameters from context, increasing flexibility while introducing malformed-call risks.

A Paradigm-Aware Analysis

The paradigm-aware taxonomy organizes the reviewed literature by architectural lineage, domain, governance focus, and publication period. It reveals domain specialization, a governance imbalance, and a shift toward neural orchestration after the rise of LLMs.

  • Taxonomy: The taxonomy classifies the complete corpus of 90 studies by primary architectural paradigm.
  • Paradigm specialization by domain: Healthcare and legal technology favor symbolic or highly constrained neural architectures, while finance leverages neural orchestration for complex analysis.
  • The governance divide: Ethics and governance research focuses overwhelmingly on neural-paradigm challenges, leaving a significant gap in governance frameworks for purely symbolic systems.
  • Temporal paradigm shift: Symbolic and hybrid cognitive architectures dominated research from 2018–2021, while neural orchestration frameworks dominated after 2022.
  • Key insights: The taxonomy identifies symbolic and hybrid architectures as dominant in safety-critical healthcare and robotics, while pure neural systems thrive in adaptive finance and education.

7 Ethical and Governance Challenges: A Paradigm-Specific Analysis

Agentic AI governance must account for differences between symbolic and neural risks rather than treating ethical challenges as monolithic. The analysis further argues that oversight should scale with agency and address the compounded complexity of hybrid systems.

  • Paradigm-specific governance: Agentic AI creates distinct ethical and governance risks across symbolic and neural paradigms, requiring paradigm-aware oversight.
  • Paradigm-specific governance: Effective governance cannot be architecturally agnostic because full explainability may be feasible for symbolic systems but technologically impossible for pure neural agents.
  • Hybrid oversight: Hybrid systems inherit both paradigms’ governance challenges, requiring audits of symbolic logic and monitoring for neural stochastic failures.
  • Accountability: Neural systems create an attribution gap because their diffuse, stochastic operation challenges legal frameworks based on direct causation and intent.
  • Human-AI collaboration: Human oversight differs by paradigm: symbolic agents require checking logical steps, whereas neural agents require steering context and interpreting opaque outputs.
  • Agency-based policy: Policy frameworks should distinguish assistive, shared, and delegated agency, scaling ethical principles and legal accountability with autonomy.

8 Research Gaps: A Paradigm-Specific Roadmap

Research gaps in Agentic AI are paradigm-specific rather than uniform: tools, benchmarks, governance, and success criteria must distinguish symbolic and neural systems while enabling their integration.

  • Paradigm-aware methodologies are needed because tools, benchmarks, and success criteria developed for one paradigm can be irrelevant or misapplied to the other.
  • Neuro-symbolic architectures are a promising response to the reasoning-and-adaptability gap, combining neural pattern recognition with symbolic logical guidance and constraints.
  • Evaluation should use separate rigorous suites that test paradigm-specific failures, including logic bombs and edge-case reasoning for symbolic systems and prompt injection resilience and output stability for neural systems.
  • Governance requires tailored frameworks: symbolic systems need formal verifiability and logic auditing, whereas neural systems require statistical or training-based alignment and audits of data and prompts.
  • A dual-track roadmap should deepen understanding of each paradigm while developing integrated architectures and standards for reliable, trustworthy agentic systems.

9 Future Directions: The Path to Hybrid Intelligence

Future Agentic AI development is oriented toward architectural integration: hybrid systems should combine symbolic reliability with neural adaptability, supported by specialized ecosystems and dual-track governance.

  • Future Directions: The Path to Hybrid Intelligence: Architectural integration is the overriding future direction, combining the reliability of symbolic reasoning with the adaptability of neural generation.
  • Future Directions: The Path to Hybrid Intelligence: Neuro-symbolic systems couple neural perception and pattern recognition with symbolic reasoning and constraint checking to address the brittleness and opacity of isolated paradigms.
  • Future Directions: The Path to Hybrid Intelligence: Future multi-agent ecosystems will assign specialized roles to neural and symbolic agents that communicate through standardized protocols.
  • Future Directions: The Path to Hybrid Intelligence: Governance should advance formal methods for symbolic verifiability alongside statistical, training-based methods for neural alignment, then combine both for hybrid agents.
  • Future Directions: The Path to Hybrid Intelligence: Neural agents can manage real-time IoT sensor data while symbolic modules ensure resulting decisions are safe and compliant.
  • Future Directions: The Path to Hybrid Intelligence: The future is a paradigm-aware synthesis integrating ethics, interdisciplinary methods, and infrastructure-aware governance to create intelligence that is adaptable and reliable.

10 Conclusion

The review establishes two distinct Agentic AI lineages with different operational mechanics and domain strengths, and concludes that progress depends on their strategic integration into trustworthy hybrid systems.

  • The review introduces and validates Symbolic/Classical and Neural/Generative lineages with fundamentally different operational mechanics, strengths, and limitations.
  • Symbolic systems suit safety, verifiability, and explicit logic, while neural systems suit adaptability, pattern recognition, and unstructured data.
  • Governance must be paradigm-specific: symbolic accountability involves auditing logic, while neural accountability involves auditing training data and prompts.
  • The review identifies strategic integration into neuro-symbolic architectures as the most productive path, leveraging symbolic reliability and neural adaptability.
  • The dual-paradigm framework provides vocabulary and taxonomy for classifying systems, evaluating capabilities and risks, and making informed design choices.
  • Agentic AI’s success is sociotechnical, requiring hybrid intelligence that is both powerful and trustworthy for scientific, service, and infrastructure applications.
Loading 2510.25445v1…