Aussie AI

Agent AI Architectures

  • Last Updated 27 February, 2025
  • by David Spuler, Ph.D.

Agents. Ah, yes, the final phase of AI. You tell your AI assistant to go away and book you a vacation, and then you're done. Sorted! (What could go wrong?)

Agent architectures are a very powerful extension of LLMs, especially if you use the trendy "agentic" terminology. There are literally dozens of startups building them, but no-one really agrees what they are. Here's one way to look at it using computer programmer vernacular:

  • Read-only agents — go away and research, coming back to give you more detailed answers.
  • Read-write agents — they go away and research it, book your vacation, and then tell you about it.

Actually, to further confuse issues, there are a few agent-like architectures already available and in common usage:

  • "Plug-ins" (such as for OpenAI's ChatGPT) — access third-party data sources in your LLM queries, which is a RAG-like architecture whereby the AI engine gets more data from the integrated data source, and then uses the LLM to complete the answer.
  • "Tools" — your AI actually launches some executable tools behind-the-scenes, such as a clock to answer time-related queries, a calculator for arithmetic, and so on. The LLM has to be trained to know when to launch a tool, and which ones to use for what queries.

There's also a time-based categorization possible:

  • Interactive agents — launch now, do their best to handle it quickly, and then report on what they found (or what they did). For example, you tell your agent to send a text message or an email.
  • Launched agents — the idea is longer-running agents, such as one that goes off and books your concert tickets.
  • Scheduled, Autonomous, or Perpetual agents — the idea is for an agent to sit there, always watching, and then report on this, or get triggered by something (e.g. it watches a newsfeed and summarizes them for you each morning; or an agent watches stock prices, and then trades for you as your own personal HFT).

Some of the other practical issues involved in running an agent include:

  • Approval-needed versus unattended execution. (It's all about trust, or lack thereof.)
  • Follow-up questions, whereby the AI determines what you're asking, but then reverses the power balance in the relationship by asking for clarification of your wishes ("window or aisle?").
  • Personalization — it really should already know that you want a window seat.
  • Context — the agent would ideally know your geo location, or what's on your screen, or where you're standing (e.g. it really should look out through your camera and know that you're in the bathroom, as everyone would want).

Related areas of LLM research include:

Survey Papers on AI Agents

Survey papers on agents:

  • Z. Xi, W. Chen, X. Guo, W. He, Y. Ding, B. Hong, M. Zhang, J. Wang, S. Jin, E. Zhou et al., “The rise and potential of large language model based agents: A survey,” arXiv preprint arXiv:2309.07864, 2023. https://arxiv.org/abs/2309.07864
  • T. Guo, X. Chen, Y. Wang, R. Chang, S. Pei, N. V. Chawla, O. Wiest, and X. Zhang, “Large language model based multi agents: A survey of progress and challenges,” arXiv preprint arXiv:2402.01680, 2024. https://arxiv.org/abs/2402.01680
  • Junwei Liu, Kaixin Wang, Yixuan Chen, Xin Peng, Zhenpeng Chen, Lingming Zhang, Yiling Lou, 4 Sep 2024, Large Language Model-Based Agents for Software Engineering: A Survey, https://arxiv.org/abs/2409.02977 Project: https://github.com/FudanSELab/Agent4SE-Paper-List
  • Yanxian Huang, Wanjun Zhong, Ensheng Shi, Min Yang, Jiachi Chen, Hui Li, Yuchi Ma, Qianxiang Wang, Zibin Zheng, Yanlin Wang, 13 Sep 2024, Agents in Software Engineering: Survey, Landscape, and Vision, https://arxiv.org/abs/2409.09030 https://github.com/DeepSoftwareAnalytics/Awesome-Agent4SE
  • Biao Wu, Yanda Li, Meng Fang, Zirui Song, Zhiwei Zhang, Yunchao Wei, Ling Chen, 4 Nov 2024, Foundations and Recent Trends in Multimodal Mobile Agents: A Survey, https://arxiv.org/abs/2411.02006 https://github.com/aialt/awesome-mobile-agents
  • Shuai Wang, Weiwen Liu, Jingxuan Chen, Weinan Gan, Xingshan Zeng, Shuai Yu, Xinlong Hao, Kun Shao, Yasheng Wang, Ruiming Tang, 7 Nov 2024, GUI Agents with Foundation Models: A Comprehensive Survey, https://arxiv.org/abs/2411.04890
  • Chris Sypherd, Vaishak Belle, 5 Dec 2024, Practical Considerations for Agentic LLM Systems, https://arxiv.org/abs/2412.04093
  • Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, Xintong Li, Jing Shi, Hongjie Chen, Viet Dac Lai, Zhouhang Xie, Sungchul Kim, Ruiyi Zhang, Tong Yu, Mehrab Tanjim, Nesreen K. Ahmed, Puneet Mathur, Seunghyun Yoon, Lina Yao, Branislav Kveton, Thien Huu Nguyen, Trung Bui, Tianyi Zhou, Ryan A. Rossi, Franck Dernoncourt, 18 Dec 2024, GUI Agents: A Survey, https://arxiv.org/abs/2412.13501
  • Shuaihang Chen, Yuanxing Liu, Wei Han, Weinan Zhang, Ting Liu, 23 Dec 2024. A Survey on Multi-Generative Agent System: Recent Advances and New Frontiers. https://arxiv.org/abs/2412.17481
  • Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei, 15 Jan 2025, Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG, https://arxiv.org/abs/2501.09136
  • Xinzhe Li, Jan 2025, A Review of Prominent Paradigms for LLM-Based Agents: Tool Use (Including RAG), Planning, and Feedback Learning, Proceedings of the 31st International Conference on Computational Linguistics, pages 9760–9779, January 19–24, 2025. ©2025 Association for Computational Linguistics, https://aclanthology.org/2025.coling-main.652.pdf https://github.com/xinzhel/LLM-Agent-Survey

Building AI Agents

Research papers on how to build AI agents:

Research on AI Agent Architectures

There are plenty of research papers on agent extensions to AI.

Multi-Agent Architectures

Research papers on multi-agent architectures:

Agentic Workflow

Agentic Architectures

Agent Frameworks and Platforms

Agent frameworks, or agent platforms, are overall LLM infrastructure to create and run LLM agents. Various commercial and open-source agent platforms have gained prominence, and there is much more to come in this space.

More AI Research

Read more about: