Agentic AI Essentials

Lesson Overview

Welcome to the first lesson of Agentic AI Essentials. In this course, you’ll learn how modern AI “agents” – powered by Large Language Models (LLMs) – can tackle challenges like customer support, research assistance, and workflow automation. By blending advanced language understanding (from tools like ChatGPT, GPT-4, and Gemini) with decision-making processes, these agents can perform tasks autonomously, reduce operational overhead, and accelerate innovation within organizations.


What You’ll Learn in This Lesson

  1. Defining Agentic AI: The “agent” concept and why autonomy matters.
  2. LLMs Demystified: An overview of GPT-4, ChatGPT, Gemini, and similar advanced language models.
  3. Beyond Automation: How agentic AI goes further than traditional rule-based systems.
  4. Business Impact: Early success stories and examples from real companies.
  5. Key Terms & Concepts: Get comfortable with prompts, tokens, embeddings, RAGs, and more.

1. What Is Agentic AI?

Agentic AI Architecture
This diagram illustrates the key components and workflow of an Agentic AI system, showcasing how autonomous agents interact within an environment to achieve specific goals.

Agentic AI describes AI systems that can handle tasks and make decisions with minimal human supervision. Think of an AI that doesn’t wait for every command but instead proactively suggests actions or corrects its course when it encounters an obstacle.

“Agentic AI = Autonomy + Adaptability + Goal-Driven Behavior.”

Instead of following a rigid set of instructions, an agentic system:

  • Receives an overarching goal or objective.
  • Observes relevant data or context (like chat history or document repositories).
  • Takes actions based on learned patterns, “thinking” through the steps.
  • Iterates or refines its approach if new information or errors arise.

This shift from rule-based to goal-based operation has wide-ranging implications, from chatbots that autonomously schedule appointments to research assistants that automatically compile summaries.

Here’s Andrew Ng explaining how agentic workflows will revolutionise AI.


2. A Layman’s Breakdown of LLMs

Transformer Model Architecture
A visual representation of the Transformer architecture, fundamental to the development of LLMs, highlighting the encoder-decoder structure and attention mechanisms.

Large Language Models (LLMs) are AI systems trained on enormous text corpora—imagine billions of documents, articles, and websites. Examples include:

These models “learn” language patterns, grammar, facts, and even common-sense reasoning. When you prompt an LLM with a question or request, it draws on this wealth of knowledge to predict the most appropriate responses.


3. Introducing RAGs (Retrieval-Augmented Generation)


This diagram illustrates how Retrieval-Augmented Generation combines external data retrieval with LLM-generated responses.

What is RAG?
Retrieval-Augmented Generation (RAG) is a technique that enhances the capabilities of LLMs by integrating them with real-time or external data retrieval systems. Instead of relying solely on pre-trained knowledge, the model retrieves relevant, up-to-date information from external sources—such as a database, knowledge graph, or the web—before generating a response.

Why RAG Matters for Agentic AI:

  • Contextual Relevance: RAG allows AI agents to provide responses based on the latest and most specific data, such as a company’s internal documentation or current market trends.
  • Overcoming Model Limitations: LLMs trained on static datasets may lack up-to-date or domain-specific information. RAG bridges this gap.
  • Enhanced Accuracy: By grounding outputs in real-time facts or verified documents, RAG reduces the risk of “hallucinations” (AI generating inaccurate or fabricated content).

How It Works:

  1. Querying External Data Sources: The AI agent sends a query to a connected database or search engine.
  2. Retrieving Relevant Information: The system fetches relevant data and feeds it back to the LLM.
  3. Response Generation: The LLM uses both the retrieved data and its own knowledge to craft a more accurate and contextually rich response.

Example Use Case:
A customer service AI assistant might use RAG to pull real-time shipping information from a logistics database, enabling it to answer customer queries about order statuses with up-to-date details.


4. Examples of AI Agents in Action

Multi-Agent System Architecture
This image depicts a multi-agent system architecture, where multiple agents collaborate to achieve a common goal within a business or physical environment context.

Here are some real-world implementations of LLM-driven agents making a difference:

  • Smart Helpdesk:
    A global IT services firm deployed an AI assistant that automatically triages incoming support tickets, suggests troubleshooting steps, and even routes complex issues to the correct department. (See IBM Watson Assistant case studies for more details.)
  • Market Research Bot:
    A marketing agency uses a GPT-4-based agent to scrape news sites, monitor social media trends, and generate weekly reports on consumer sentiment. This reduces hours of manual research.
  • RAG-Powered Knowledge Assistants:
    A legal AI agent combines GPT-4 with RAG techniques to retrieve laws and precedents from databases, helping legal teams draft arguments based on the latest legal context.

5. Why Are LLM-Powered Agents So Popular Now?

Overview of Large Language Models
An illustration detailing the architecture of Large Language Models, including layers such as feed-forward layers, attention mechanisms, and positional encoding.

The surge in popularity can be traced to improvements in computing power, the availability of massive training datasets, and critical AI research breakthroughs. For instance, Google’s Gemini integrates multimodal capabilities (text, image, and video understanding) into a single model, enhancing flexibility.

From a business perspective, managers and leaders see these benefits:

  • Cost Savings: Automation of tasks such as document processing, data analysis, or repetitive inquiries.
  • 24/7 Availability: Agents never need sleep, ensuring continuous customer support or internal assistance.
  • Scalability: One AI agent can handle the workload of multiple human operators in many routine tasks.
  • Enhanced Insights: LLMs and RAGs can sift through massive datasets (financial reports, research papers, social media feeds) and provide actionable summaries or recommendations.

Key Takeaways

  • Agentic AI represents a paradigm shift from static, rule-based systems toward adaptive, autonomous solutions that can handle evolving tasks.
  • LLMs like ChatGPT, GPT-4, Claude, and Gemini are the “brains” behind modern AI agents, empowering them to understand context and produce human-like outputs.
  • RAGs (Retrieval-Augmented Generation) allow agents to retrieve live, specific data for enhanced accuracy and relevance.
  • The Business Impact is already visible: from intelligent chatbots that reduce call-center volume to agents that comb through documents and data at scale.


Frequently asked questions

What is the difference between a standard LLM and an agentic AI system?

A standard LLM processes text to generate responses, while an agentic AI system uses that model to plan and execute multi-step tasks. Agents can use tools, browse the web, and make decisions without constant human input. This shift moves AI from passive chat to active problem solving.

How does the Model Context Protocol (MCP) work for AI agents?

MCP acts as a universal connector between AI models and external data sources. It standardises how agents access files, databases, and APIs, removing the need for custom integrations. Think of it as a USB-C port for AI, allowing any compatible agent to plug into new tools instantly.

Can non-technical managers build AI agents without writing code?

Yes, platforms like Zapier and ChatGPT allow managers to create functional agents using visual interfaces. You can connect existing business apps and define logic through natural language prompts. This approach removes the need for Python or complex engineering skills, making automation accessible to operational teams.

What are the main ethical risks of deploying autonomous AI agents?

The primary risks include hallucinations leading to incorrect actions and lack of transparency in decision making. Agents may also bypass security protocols if not properly constrained. Organisations must implement strict governance, audit logs, and human oversight to ensure agents operate within defined ethical boundaries.

How much does it cost to run an AI agent prototype?

Costs vary significantly based on usage and complexity. A simple prototype using Zapier and a standard ChatGPT subscription might cost between twenty and fifty pounds per month. More complex agents requiring custom API calls or higher token limits can exceed one hundred pounds monthly, depending on the volume of tasks processed.

Web3 & RWA Program • The Tesseract Academy

Tokenisation 101 for Entrepreneurs & Fund Managers

Comprehensive framework on Real-World Asset (RWA) tokenisation, SPV structures, and regulatory compliance.