Fundamentals of Building Autonomous LLM Agents

Abstract. This paper reviews the architecture and implementation methods of agents powered by large language models (LLMs). Motivated by the limitations of traditional LLMs in real-world tasks, the research aims to explore patterns to develop “agentic” LLMs that can automate complex tasks and bridge the performance gap with human capabilities. Key components include a perception system that converts environmental percepts into meaningful representations; a reasoning system that formulates plans, adapts to feedback, and evaluates actions through different techniques like Chain-of-Thought and Tree-of-Thought; a memory system that retains knowledge through both short-term and long-term mechanisms; and an execution system that translates internal decisions into concrete actions. This paper shows how integrating these systems leads to more capable and generalized software bots that mimic human cognitive processes for autonomous and intelligent behavior.

Introduction. Artificial intelligence (AI) is a powerful technology that is transforming cognitive automation and fundamentally reshaping the way tasks are performed [13,14,37]. Today, one can develop remarkable systems without the need to write complex algorithms or master low-level code. We are closer than ever to realizing the idea that “if you can think it, you can build it.” Instead of relying solely on programming skills, what increasingly matters is understanding how a human would reason through a problem, since LLM agents can learn and mimic human problem solving by externalizing intermediate reasoning and refining it through self-feedback [26,38,49,58,60,65,66]. LLM agents represent a new paradigm that breaks traditional barriers. They enable the execution of tasks that were previously costly, time-consuming, or even infeasible. More than tools, agents act as collaborators, assisting humans in dynamic environments and automating decision-making in critical systems. However, this transformation is still in its early stages.

Discussion / Conclusion. While our review sheds light on the foundational elements of intelligent LLM agents, several limitations warrant consideration. Firstly, these agents currently fail at certain operations that humans can easily perform, largely due to a lack of sufficient experience interacting in specific environments. Teaching these experiences to LLMs is exceptionally costly, often requiring extensive fine-tuning. This challenge is compounded by the fact that many advanced models are closedsource, making it difficult to fine-tune this models. Moreover, acquiring the necessary data for targeted training is also time-consuming. Secondly, while LLMs excel at generating and understanding text, their ability to generate precise actions in the real world or within graphical user interfaces (GUIs) remains limited. Thirdly, despite advancements, visual perception in these agents is not yet as robust as required, with many mistakes stemming from an incomplete or inaccurate understanding of the environment. The review presented in this paper has significant implications for the future of artificial intelligence.