Agentic AI Systems: A Review of Multi-Agent Reasoning, Trust, Orchestration, and Autonomous Data Science

Abstract

Agentic artificial intelligence (AI) — systems of autonomous, goal-directed software agents that plan, act, and coordinate with reduced human oversight — is moving from research prototypes toward applied data-science, perception, and decision-support pipelines. Yet the literature remains fragmented across largely disconnected strands: trust-aware orchestration of multimodal agents for visual classification, human adoption of autonomous AI agents, automated machine-learning (AutoML) pipelines for tabular data, and fully autonomous multi-agent systems for end-to-end data engineering. This review synthesizes four recent studies spanning these strands, together with the broader literature on multi-agent orchestration, retrieval-augmented generation (RAG), large-language-model (LLM)-based knowledge-graph construction, and explainable AI, to build a unified picture of agentic AI as applied to intelligent data science. We propose a four-part taxonomy — perception-level trust orchestration, human adoption of agentic systems, automated feature/model pipelines, and autonomous data-engineering agents — and compare the reviewed studies on automation level, reasoning mechanism, knowledge representation, retrieval use, trust treatment, and evaluation methodology. The synthesis shows that individual components of an agentic data-science stack (confidence calibration, retrieval-grounded re-evaluation, automated feature synthesis, autonomous pipeline repair) have each demonstrated measurable gains in isolation, but no reviewed system couples them with a persistent, multimodal knowledge-graph layer, and adoption research indicates that user trust is not the dominant driver of usage intention — a finding with direct design implications. We identify recurring gaps in provenance, explainability, benchmark standardization, and graph-grounded reasoning, and outline an integrated conceptual framework for the reviewed research landscape, positioning closer integration of agentic data science with knowledge-graph representation as an open research trajectory rather than a solved problem.

Introduction

This paper reviews recent research on agentic AI, focusing on how autonomous and multi-agent AI systems can become more trustworthy, useful, and capable of reasoning over complex data. It identifies a major gap: research on trust-aware agents, human adoption, AutoML, autonomous data engineering, and knowledge graphs has largely developed in separate communities without a unified framework.

Main purpose

The review aims to:

• Compare four representative studies on agentic AI.

• Examine how trust, autonomy, automation, and human adoption are addressed.

• Determine whether existing AutoML and data-engineering systems truly demonstrate agentic behavior.

• Identify gaps in integrating agentic AI with structured and multimodal knowledge, especially knowledge graphs.

• Propose directions for future research.

Methodology

The study uses a structured narrative/scoping review rather than a full systematic review. Four seed papers are analyzed and supplemented with research from sources such as IEEE Xplore, arXiv, and ACM Digital Library. Literature was searched using terms related to agentic AI, multi-agent orchestration, trust calibration, AutoML, autonomous data engineering, knowledge graphs, and GraphRAG.

The four studies represent:

1. Trust-aware multi-agent orchestration for visual classification.

2. Human adoption and acceptance of autonomous AI agents.

3. Automated data science/AutoML.

4. Autonomous multi-agent data engineering.

Key concepts

The paper distinguishes AI agents from agentic AI. Traditional AI agents generally perform specific tasks, whereas agentic AI involves greater autonomy, task decomposition, multiple specialized agents, orchestration, memory, and reduced human supervision.

Trust is examined in two different ways:

• Technical trust: measured through calibration metrics such as Expected Calibration Error (ECE), Overconfidence Ratio, and consistency.

• Human trust: treated as a psychological factor influencing users' willingness to adopt AI systems.

The review also discusses RAG, where AI retrieves external evidence to improve its responses, and GraphRAG, which retrieves information from knowledge graphs to support more structured and multi-hop reasoning.

Four research categories

The reviewed literature is organized into four major categories:

1. Trust-aware multi-agent orchestration – multiple specialized agents are coordinated by an orchestrator that evaluates confidence and triggers additional verification.

2. Human adoption of agentic AI – studies examine factors such as usefulness, risk, trust, and perceived humanness that influence users' willingness to adopt AI agents.

3. Automated machine learning (AutoML) – systems automate preprocessing, feature engineering, model selection, hyperparameter tuning, and interpretation.

4. Autonomous data engineering – multi-agent systems automate data ingestion, schema management, transformation, validation, and deployment.

Major findings

The trust-aware visual-classification study showed that combining multiple vision-language agents, an orchestrator, trust calibration, and image-based retrieval substantially improved accuracy—from 48.13% in a confidence-aware zero-shot setting to 85.63% with trust-aware RAG, although supervised fine-tuning achieved 97.50%.

However, the study revealed an important weakness: when multiple agents confidently agree on the same incorrect answer, the orchestrator may lack a dissenting signal needed to detect the error.

The review also emphasizes that technical reliability and human trust are different concepts. An AI system can be technically well calibrated but still not be accepted by users, while users may trust systems that are not objectively reliable.

Major research gap

A central finding is that the four studies generally work with images, text, or tabular data, but do not adequately integrate agentic reasoning with knowledge graphs and multimodal structured knowledge.

The paper therefore identifies an opportunity to combine:
multi-agent reasoning + trustworthy orchestration + automated data science/data engineering + knowledge graphs + GraphRAG + explainability.

Such an integrated system could allow AI agents to autonomously process data, construct and query structured knowledge, verify their reasoning, and provide more traceable and reliable conclusions.

Conclusion

This review analyzed and synthesized four papers spanning trust-aware multi-agent orchestration for visual classification [1], human adoption of autonomous AI agents [2], an interpretable AutoML pipeline [3], and an autonomous end-to-end data-engineering system [4], situating them within the broader literature on agentic AI, retrieval-augmented and graph-based generation, LLM-driven knowledge-graph construction, and explainable AI. The comparative analysis shows that current agentic-AI research has independently matured several component capabilities — confidence-calibrated orchestration, retrieval-grounded re-evaluation, automated feature engineering, and autonomous pipeline governance — while leaving their integration with structured, multimodal knowledge representation, and their alignment with empirically grounded user-adoption principles, largely unaddressed. These gaps motivate, as a conceptual research direction rather than a demonstrated result, closer integration of autonomous agentic data science with multimodal knowledge-graph generation. Future empirical work should validate the proposed conceptual framework, formalize systematic review protocols, and develop shared benchmarks that allow the field to compare agentic data-science systems on common ground.

References

[1] K. I. Roumeliotis, R. Sapkota, M. Karkee, and N. D. Tselikas, \"Agentic AI With Orchestrator-Agent Trust: A Modular Visual Classification Framework With Trust-Aware Orchestration and RAG-Based Reasoning,\" IEEE Access, vol. 14, pp. 26965–26982, 2026. [2] M. Hasselwander and O. Lah, \"Agentic AI Arrives: How Gen Z Adopts Autonomous AI Agents,\" IEEE Access, vol. 14, pp. 27083–27090, 2026. [3] I. Hasan and M. Tausif, \"Designing an Interpretable and Efficient AutoML Pipeline for Enhanced Data Analytics,\" in Proc. 3rd Int. Conf. Self Sustainable Artif. Intell. Syst. (ICSSAS), 2025, pp. 911–916. [4] K. R. Chirumamilla, \"Autonomous AI System for End-to-End Data Engineering,\" Int. J. Intell. Syst. Appl. Eng. (IJISAE), vol. 12, no. 13s, pp. 790–801, 2024. [5] R. Sapkota, K. I. Roumeliotis, and M. Karkee, \"AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges,\" Inf. Fusion, vol. 126, Art. no. 103599, Feb. 2026. [6] S. Murugesan, \"The Rise of Agentic AI: Implications, Concerns, and the Path Forward,\" IEEE Intell. Syst., vol. 40, no. 2, pp. 8–14, Mar. 2025. [7] Y. Zhu, X. Wang, J. Chen, S. Qiao, Y. Ou, Y. Yao, S. Deng, H. Chen, and N. Zhang, \"LLMs for Knowledge Graph Construction and Reasoning: Recent Capabilities and Future Opportunities,\" arXiv:2305.13168, 2024. [8] B. Peng, Y. Zhu, Y. Liu, X. Bo, H. Shi, C. Hong, Y. Zhang, and S. Tang, \"Graph Retrieval-Augmented Generation: A Survey,\" arXiv:2408.08921, 2024. [9] D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, D. Metropolitansky, R. O. Ness, and J. Larson, \"From Local to Global: A Graph RAG Approach to Query-Focused Summarization,\" arXiv:2404.16130, 2025. [10] Q. Zhang, S. Chen, Y. Bei, Z. Yuan, H. Zhou, Z. Hong, J. Dong, H. Chen, Y. Chang, and X. Huang, \"A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models,\" arXiv:2501.13958, 2025. [11] H. Han, Y. Wang, H. Shomer, K. Guo, J. Ding, Y. Lei, M. Halappanavar, R. A. Rossi, S. Mukherjee, X. Tang, Q. He, Z. Hua, B. Long, T. Zhao, N. Shah, A. Javari, Y. Xia, and J. Tang, \"Retrieval-Augmented Generation with Graphs (GraphRAG),\" arXiv:2501.00309, 2024. [12] M. Feurer, K. Eggensperger, S. Falkner, M. Lindauer, and F. Hutter, \"Auto-Sklearn 2.0: Hands-Free AutoML via Meta-Learning,\" J. Mach. Learn. Res., vol. 23, no. 1, Art. no. 261, Jan. 2022. [13] J. M. Kanter and K. Veeramachaneni, \"Deep Feature Synthesis: Towards Automating Data Science Endeavors,\" in Proc. IEEE Int. Conf. Data Sci. Adv. Analytics (DSAA), 2015, pp. 1–10. [14] M. T. Ribeiro, S. Singh, and C. Guestrin, \"\'Why Should I Trust You?\' Explaining the Predictions of Any Classifier,\" in Proc. ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2016, pp. 1135–1144. [15] S. Afroogh, A. Akbari, E. Malone, M. Kargar, and H. Alambeigi, \"Trust in AI: Progress, Challenges, and Future Directions,\" Humanities Soc. Sci. Commun., vol. 11, no. 1, pp. 1–30, Nov. 2024. [16] L. Longo et al., \"Explainable Artificial Intelligence (XAI) 2.0: A Manifesto of Open Challenges and Interdisciplinary Research Directions,\" Inf. Fusion, vol. 106, Art. no. 102301, Jun. 2024. [17] S. Ali et al., \"Explainable Artificial Intelligence (XAI): What We Know and What Is Left to Attain Trustworthy Artificial Intelligence,\" Inf. Fusion, vol. 99, Art. no. 101805, Nov. 2023. [18] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, ?. Kaiser, and I. Polosukhin, \"Attention Is All You Need,\" in Adv. Neural Inf. Process. Syst. (NeurIPS), 2017.

Copyright

Copyright © 2026 Subina S B, Abhijith B, Pranav Edwin, Sania Jackson , Geethu Raj. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.