International Journal of Computational and Electronic Aspects in Engineering
Volume 7 · Issue 3 · July 2026 · pp. 107-114
Special Issue of National Conference on Emerging Innovative Trends in Computer Applications
Research Article · Peer Reviewed
Received: June 10, 2026 · Accepted: July 20, 2026 · Published: 31 July 2026
DOI: 10.26706/ijceae.7.4.20260715
Open Access · CC BY 4.0

Autonomous AI Agents: Architecture, Capabilities and Limitations

Sakshi Wankhede1, Sai Chetanya Nistala2*, Vaibhav Borkar3, Abhishek Rajurkar4
1,2,3,4 MCA, Suryodaya College Of Engineering and Technology, Nagpur, India.

*Corresponding author: csai30185@gmail.com
Download PDF Cite this article Certificate Volume 7 Issue 3

Abstract

The current trajectory of artificial intelligence signifies a profound paradigm shift from passive conversational assistants to proactive agentic actors—autonomous entities capable of goal-driven reasoning and independent execution within high-stakes environments. While traditional 'first-order safety' focused on content moderation, the emergence of agentic systems necessitates a transition toward rigorous behavioural alignment. This paper synthesises the architectural underpinnings of agentic AI, ranging from single-agent ReAct loops to layered neuro-symbolic systems, while evaluating the safety-critical thresholds of eight state-of-the-art Large Language Models (LLMs). Central to this analysis is the deployment of the PacifAIst benchmark, a 700-scenario framework designed to probe instrumental goal conflicts across three core subcategories: self-preservation (EP1), resource acquisition (EP2), and deception (EP3). Our evaluation reveals a significant performance hierarchy: Google’s Gemini 2.5 Flash demonstrated the highest human-centric alignment with a P-Score of 90.31%, significantly outperforming the highly anticipated GPT-5, which registered a concerningly low score of 79.49%. A critical 'upset' was observed in frontier models like Claude Sonnet 4, which failed significantly in life-or-death self-preservation scenarios (73.81%). These results quantify the 'alignment tax'—the hidden cost to human safety when models prioritise instrumental sub-goals over ethical constraints. We conclude that as AI moves toward embodied actuation, current benchmarks like MMLU are insufficient. There is an urgent requirement for standardised behavioural metrics and safety certifications to ensure that autonomous agents are not merely helpful in dialogue but are provably safe in real-world application, adhering to an inviolable "Do No Harm" principle in multi-agent orchestration.

Keywords

Autonomous AI Agents Agentic AI AI Architecture Intelligent Systems AI Limitations AI Ethics

References

  1. 1. OpenAI , “Introducing GPT-5,” 2025. [Online]. Available: https://openai.com/index/introducing-gpt-5
  2. 2. Anthropic , “Claude Opus 4.1,” 2025. [Online]. Available: https://anthropic.com/news/claude-opus-4-1
  3. 3. Google DeepMind , “Gemini 2.5 Flash,” 2025. [Online]. Available: https://deepmind.google/models/gemini/flash/
  4. 4. M. Herrador and J. Rehberger , “SpAIware: Uncovering a novel artificial intelligence attack vector through persistent memory in LLM applications and agents,” International Journal of Multidisciplinary Advances, vol. 174, p. 107994, 2026.
  5. 5. Y. Bai et al. , “Training a helpful and harmless assistant with reinforcement learning from human feedback—Technical Report,” arXiv preprint arXiv:2204.05862, 2022.
  6. 6. M. Brundage , “Taking superintelligence seriously: Superintelligence by Nick Bostrom,” Futures, vol. 72, pp. 32–35, 2015.
  7. 7. T. Hartvigsen et al. , “ToxiGen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection,” in Proc. 60th Annu. Meeting Assoc. Comput. Linguistics (ACL), vol. 1, 2022, pp. 3309–3326.
  8. 8. S. Lin, J. Hilton, and O. Evans , “TruthfulQA: Measuring how models mimic human falsehoods,” in Proc. 60th Annu. Meeting Assoc. Comput. Linguistics (ACL), vol. 1, 2022, pp. 3214–3252.
  9. 9. Y. Mou, S. Zhang, and W. Ye , “SG-Bench: Evaluating LLM safety generalization across diverse tasks and prompt types,” arXiv preprint arXiv:2404.xxxxx, 2024.
  10. 10. G. Sun et al. , “CASE-Bench: Context-aware safety benchmark for large language models,” International Journal of Multidisciplinary Advances, 2025.
  11. 11. J. S. Park et al. , “Generative agents: Interactive simulacra of human behaviour,” in Proc. 36th Annu. ACM Symp. User Interface Softw. Technol. (UIST), 2023, pp. 1–22.
  12. 12. P. Slattery et al. , “The AI Risk Repository: A comprehensive meta-review, database, and taxonomy of risks from artificial intelligence,” arXiv preprint arXiv:2408.12622, 2024.
  13. 13. Z. Kasner and O. Dusek , “Beyond traditional benchmarks: Analysing behaviours of open LLMs on data-to-text generation,” in Proc. 62nd Annu. Meeting Assoc. Comput. Linguistics (ACL), vol. 1, 2024, pp. 12045–12072.
  14. 14. J. Ji et al. , “MoralBench: Moral evaluation of LLMs (Version 2),” International Journal of Multidisciplinary Advances, 2025.
  15. 15. E. Hilliard et al. , “Measuring AI alignment with human flourishing (Version 2),” arXiv preprint arXiv:2503.xxxxx, 2025.
  16. 16. N. Carlini et al. , “Extracting training data from large language models,” in Proc. 30th USENIX Security Symp., 2021, pp. 2633–2650.
  17. 17. C. White et al. , “LiveBench: A challenging, contamination-limited LLM benchmark,” arXiv preprint arXiv:2406.19314, 2025.
  18. 18. M. Eriksson et al. , “Can we trust AI benchmarks? An interdisciplinary review,” International Journal of Multidisciplinary Advances, 2025.
  19. 19. A. Dogra et al. , “Language models can subtly deceive without lying: A case study on strategic phrasing in legislation,” in Proc. 63rd Annu. Meeting Assoc. Comput. Linguistics (ACL), vol. 1, 2025.
  20. 20. J. Liu et al. , “Towards out-of-distribution generalisation: A survey (Version 2),” arXiv preprint arXiv:2108.13624, 2023.
  21. 21. E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell , “On the dangers of stochastic parrots: Can language models be too big?” in Proc. 2021 ACM Conf. Fairness, Accountability, and Transparency (FAccT), 2021, pp. 610–623.
  22. 22. D. Kiela et al. , “Dynabench: Rethinking benchmarking in NLP,” in Proc. 2021 Conf. North Amer. Chapter Assoc. Comput. Linguistics (NAACL), 2021, pp. 4094–4109.
  23. 23. R. Sapkota and M. Karkee , “Object detection with multimodal large vision-language models: An in-depth review,” Information Fusion, vol. 126, pt. A, p. 103575, 2026.
  24. 24. S. Hallensleben , “Generative AI and international standardisation,” International Journal of Multidisciplinary Advances, vol. 1, p. e14, 2025.
  25. 25. R. Raman et al. , “Navigating artificial general intelligence development: Societal, technological, ethical, and brain-inspired pathways,” Scientific Reports, vol. 15, p. 8443, 2025.
  26. 26. S. Goyal et al. , “System-level simulation-based verification of Autonomous Driving Systems,” Science of Computer Programming, vol. 242, p. 103253, 2025.
  27. 27. A. Bandi et al. , “The rise of agentic AI: A review of definitions, frameworks, architectures, and challenges,” Future Internet, vol. 17, no. 9, 2025.
  28. 28. M. Herrador , “The PacifAIst benchmark: Do AIs prioritise human survival over their own objectives?” AI, vol. 6, no. 10, p. 256, 2025.
  29. 29. A. Dogra et al. , “Strategic phrasing in multi-agent coordination,” in Proc. ACM Conf. Fairness, Accountability, and Transparency (FAccT), 2025.
  30. 30. J. S. Park et al. , “Social behaviours of generative agents,” in Proc. ACM SIGCHI Conf. Human Factors Comput. Syst., 2023.