Join our book community on Discord

https://packt.link/4Bbd9

Over the past two years, large language models (LLMs) have reshaped the landscape of AI. From simple prompt-based interactions to complex applications across industries, LLMs have evolved rapidly, fueled by breakthroughs in architecture, training techniques, and fine-tuning strategies. As their capabilities evolved, the shift from ChatGPT to today’s agentic systems as of April 2025 marks a natural progression, where the addition of reasoning, planning, and action-taking capabilities represented a major technological leap.This chapter explores the foundations of LLMs, how they’re built and consumed, and the differences between pre-trained and fine-tuned models. Most importantly, it sets the stage for the next leap forward: the emergence of AI agents.In this chapter, we will cover the following topics:

• Understanding foundation models and the rise of LLMs
• Latest significant breakthroughs
• Road to AI agents
• The need...

Technical requirements

You can access the complete code for this chapter in the book’s accompanying GitHub repository at https://github.com/PacktPublishing/AI-Agents-in-Practice.

Understanding foundation models and the rise of LLMs

AI has undergone a fundamental transformation thanks to the emergence of foundation models—versatile, general-purpose models that can be adapted across a wide range of tasks. Among them, LLMs have taken center stage, redefining how we interact with machines through natural language.

Latest significant breakthroughs

The field of GenAI has experienced rapid advancements over the past few years, with breakthroughs that pushed the boundaries of efficiency, adaptability, and reasoning capabilities. In the following sections, we are going to explore some of the latest techniques that significantly enhance GenAI models’ performance while reducing computational demands.

Road to AI agents

The rapid evolution of GenAI has taken us from simple automation to increasingly intelligent systems capable of reasoning, learning, and decision-making. In recent years, LLMs have revolutionized how we interact with the surrounding ecosystem, enabling more natural conversations and sophisticated problem-solving.Let’s explore the greatest milestones that have eventually fueled the rise of AI agents.

The need for an additional layer of intelligence: introducing AI agents

LLMs have demonstrated impressive capabilities in generating coherent text, answering questions, and even performing limited problem-solving. However, their fundamental design presents several limitations when it comes to real-world applications:

• Lack of long-term memory: Most LLMs operate within a fixed context window, meaning they forget previous interactions once the context limit is exceeded. This makes them incapable of learning from past experiences or maintaining continuity over time.
• No persistent goals or autonomy: LLMs respond to individual prompts but do not operate with a persistent goal-oriented mindset. They cannot proactively make decisions, self-correct, or refine their approach over time.
• Limited reasoning and multi-step execution: While LLMs can follow instructions within a single prompt, they struggle with executing multi-step workflows, handling complex decision-making, and maintaining logical...

Summary

Over the past two years, AI has undergone a profound transformation, shifting from simple LLM API calls to more sophisticated, interactive, and autonomous systems. The rapid evolution of LLMs has been marked by innovations such as RAG, fine-tuning advancements, and reasoning-focused architectures, all aimed at improving efficiency, adaptability, and cost-effectiveness.Despite these advances, LLMs alone are not enough to meet the growing demand for AI systems that can operate autonomously, make decisions, and interact meaningfully with their environments.This shift marks a pivotal moment in AI development, where the focus is no longer just on making models bigger but on making them smarter. Instead of treating AI as a passive tool that responds to isolated prompts, we are now designing agentic systems that can act, learn, and adapt to complex real-world tasks.In the next chapter, we will examine the emergence of AI agents, their primary components, and the various forms they may...