From models to agentic applications

As discussed so far, LLMs have been demonstrating remarkable fluency in natural language processing. However, as impressive as they are, they remain fundamentally reactive rather than proactive. They lack the ability to take independent actions, interact meaningfully with external systems, or autonomously achieve complex objectives.

To unlock the next phase of AI capabilities, we need to move beyond passive text generation and toward agentic AI—systems that can plan, reason, and take action to accomplish tasks with minimal human intervention. Before exploring the potential of agentic AI, it’s important to first understand the core limitations of LLMs that necessitate this evolution.

Limitations of traditional LLMs

Despite their advanced language capabilities, LLMs have inherent constraints that limit their effectiveness in real-world applications:

1. Lack of true understanding: LLMs generate human-like text by predicting the next most likely word based on statistical patterns in training data. However, they do not understand meaning in the way humans do. This leads to hallucinations—confidently stating false information as fact—and generating plausible but incorrect, misleading, or nonsensical outputs. As Bender et al. (2021) describe, LLMs function as “stochastic parrots”—repeating patterns without genuine comprehension.
2. Struggles with complex reasoning and problem-solving: While LLMs excel at retrieving and reformatting knowledge, they struggle with multi-step reasoning, logical puzzles, and mathematical problem-solving. They often fail to break down problems into sub-tasks or synthesize information across different contexts. Without explicit prompting techniques like chain-of-thought reasoning, their ability to deduce or infer remains unreliable.
3. Outdated knowledge and limited external access: LLMs are trained on static datasets and do not have real-time access to current events, dynamic databases, or live information sources. This makes them unsuitable for tasks requiring up-to-date knowledge, such as financial analysis, breaking news summaries, or scientific research requiring the latest findings.
4. No native tool use or action-taking abilities: LLMs operate in isolation—they cannot interact with APIs, retrieve live data, execute code, or modify external systems. This lack of tool integration makes them less effective in scenarios that require real-world actions, such as conducting web searches, automating workflows, or controlling software systems.
5. Bias, ethical concerns, and reliability issues: Because LLMs learn from large datasets that may contain biases, they can unintentionally reinforce ideological, social, or cultural biases. Importantly, even with open-source models, accessing and auditing the complete training data to identify and mitigate these biases remains challenging for most practitioners. Additionally, they can generate misleading or harmful information without understanding the ethical implications of their outputs.
6. Computational costs and efficiency challenges: Deploying and running LLMs at scale requires significant computational resources, making them costly and energy-intensive. Larger models can also introduce latency, slowing response times in real-time applications.

To overcome these limitations, AI systems must evolve from passive text generators into active agents that can plan, reason, and interact with their environment. This is where agentic AI comes in—integrating LLMs with tool use, decision-making mechanisms, and autonomous execution capabilities to enhance their functionality.

While frameworks like LangChain provide comprehensive solutions to LLM limitations, understanding fundamental prompt engineering techniques remains valuable. Approaches like few-shot learning, chain-of-thought, and structured prompting can significantly enhance model performance for specific tasks. Chapter 3 will cover these techniques in detail, showing how LangChain helps standardize and optimize prompting patterns while minimizing the need for custom prompt engineering in every application.

The next section explores how agentic AI extends the capabilities of traditional LLMs and unlocks new possibilities for automation, problem-solving, and intelligent decision-making.

Understanding LLM applications

LLM applications represent the bridge between raw model capability and practical business value. While LLMs possess impressive language processing abilities, they require thoughtful integration to deliver real-world solutions. These applications broadly fall into two categories: complex integrated applications and autonomous agents.

Complex integrated applications enhance human workflows by integrating LLMs into existing processes, including:

• Decision support systems that provide analysis and recommendations
• Content generation pipelines with human review
• Interactive tools that augment human capabilities
• Workflow automation with human oversight

Autonomous agents operate with minimal human intervention, further augmenting workflows through LLM integration. Examples include:

• Task automation agents that execute defined workflows
• Information gathering and analysis systems
• Multi-agent systems for complex task coordination

LangChain provides frameworks for both integrated applications and autonomous agents, offering flexible components that support various architectural choices. This book will explore both approaches, demonstrating how to build reliable, production-ready systems that match your specific requirements.

Autonomous systems of agents are potentially very powerful, and it’s therefore worthwhile exploring them a bit more.

Understanding AI agents

It is sometimes joked that AI is just a fancy word for ML, or AI is ML in a suit, as illustrated in this image; however, there’s more to it, as we’ll see.

Figure 1.1: ML in a suit. Generated by a model on replicate.com, Diffusers Stable Diffusion v2.1

An AI agent represents the bridge between raw cognitive capability and practical action. While an LLM possesses vast knowledge and processing ability, it remains fundamentally reactive without agency. AI agents transform this passive capability into active utility through structured workflows that parse requirements, analyze options, and execute actions.

Agentic AI enables autonomous systems to make decisions and act independently, with minimal human intervention. Unlike deterministic systems that follow fixed rules, agentic AI relies on patterns and likelihoods to make informed choices. It functions through a network of autonomous software components called agents, which learn from user behavior and large datasets to improve over time.

Agency in AI refers to a system’s ability to act independently to achieve goals. True agency means an AI system can perceive its environment, make decisions, act, and adapt over time by learning from interactions and feedback. The distinction between raw AI and agents parallels the difference between knowledge and expertise. Consider a brilliant researcher who understands complex theories but struggles with practical application. An agent system adds the crucial element of purposeful action, turning abstract capability into concrete results.

In the context of LLMs, agentic AI involves developing systems that act autonomously, understand context, adapt to new information, and collaborate with humans to solve complex challenges. These AI agents leverage LLMs to process information, generate responses, and execute tasks based on defined objectives.

Particularly, AI agents extend the capabilities of LLMs by integrating memory, tool use, and decision-making frameworks. These agents can:

• Retain and recall information across interactions.
• Utilize external tools, APIs, and databases.
• Plan and execute multi-step workflows.

The value of agency lies in reducing the need for constant human oversight. Instead of manually prompting an LLM for every request, an agent can proactively execute tasks, react to new data, and integrate with real-world applications.

AI agents are systems designed to act on behalf of users, leveraging LLMs alongside external tools, memory, and decision-making frameworks. The hope behind AI agents is that they can automate complex workflows, reducing human effort while increasing efficiency and accuracy. By allowing systems to act autonomously, agents promise to unlock new levels of automation in AI-driven applications. But are the hopes justified?

Despite their potential, AI agents face significant challenges:

• Reliability: Ensuring agents make correct, context-aware decisions without supervision is difficult.
• Generalization: Many agents work well in narrow domains but struggle with open-ended, multi-domain tasks.
• Lack of trust: Users must trust that agents will act responsibly, avoid unintended actions, and respect privacy constraints.
• Coordination complexity: Multi-agent systems often suffer from inefficiencies and miscommunication when executing tasks collaboratively.

Production-ready agent systems must address not just theoretical challenges but practical implementation hurdles like:

• Rate limitations and API quotas
• Token context overflow errors
• Hallucination management
• Cost optimization

LangChain and LangSmith provide robust solutions for these challenges, which we’ll explore in depth in Chapter 8 and Chapter 9. These chapters will cover how to build reliable, observable AI systems that can operate at an enterprise scale.

When developing agent-based systems, therefore, several key factors require careful consideration:

• Value generation: Agents must provide a clear utility that outweighs their costs in terms of setup, maintenance, and necessary human oversight. This often means starting with well-defined, high-value tasks where automation can demonstrably improve outcomes.
• Trust and safety: As agents take on more responsibility, establishing and maintaining user trust becomes crucial. This encompasses both technical reliability and transparent operation that allows users to understand and predict agent behavior.
• Standardization: As the agent ecosystem grows, standardized interfaces and protocols become essential for interoperability. This parallels the development of web standards that enabled the growth of internet applications.

While early AI systems focused on pattern matching and predefined templates, modern AI agents demonstrate emergent capabilities such as reasoning, problem-solving, and long-term planning. Today’s AI agents integrate LLMs with interactive environments, enabling them to function autonomously in complex domains.

The development of agent-based AI is a natural progression from statistical models to deep learning and now to reasoning-based systems. Modern AI agents leverage multimodal capabilities, reinforcement learning, and memory-augmented architectures to adapt to diverse tasks. This evolution marks a shift from predictive models to truly autonomous systems capable of dynamic decision-making.

Looking ahead, AI agents will continue to refine their ability to reason, plan, and act within structured and unstructured environments. The rise of open-weight models, combined with advances in agent-based AI, will likely drive the next wave of innovations in AI, expanding its applications across science, engineering, and everyday life.

With frameworks like LangChain, developers can build complex and agentic structured systems that overcome the limitations of raw LLMs. It offers built-in solutions for memory management, tool integration, and multi-step reasoning that align with the ecosystem model presented here. In the next section we will explore how LangChain facilitates the development of production-ready AI agents.