Lesson 6

AI Agents, Hallucinations & What's Next

AI is evolving fast. In this final lesson, we'll explore AI agents that can take action on your behalf, understand why AI sometimes makes things up, look at multimodal AI, and talk about why learning how to learn with AI is the real superpower.

After this lesson, you'll be able to:

  • ✓ Distinguish between AI chatbots and AI agents
  • ✓ Explain why LLMs hallucinate and how to reduce it
  • ✓ Describe what multimodal AI is and why it matters

What Are AI Agents?

AI agents are AI systems that can take actions autonomously — not just generate text, but use tools, write code, browse the web, make decisions, and complete multi-step tasks.

The key difference from chatbots: a chatbot responds to one message at a time. An agent plans, executes, observes results, and adapts. It can break a complex goal into steps, use different tools for each step, and adjust its approach based on what it finds along the way.

The Agent Loop

AI agent loop diagram An agent cycles through four steps — plan, act, observe, adapt — and repeats until the goal is achieved. Goal repeat until done Plan Act Observe Adapt

Examples of AI Agents

•

Claude Code — an AI agent that can read your codebase, write code, run terminal commands, and fix bugs across multiple files

•

Coding assistants — agents that can navigate repositories, run tests, and submit pull requests

•

Research agents — agents that can search the web, read documents, synthesize findings, and produce reports

Analogy: A chatbot is like texting a friend for advice. An AI agent is like hiring an assistant who goes and does the work for you.

Hallucination

A hallucination is when an LLM generates information that sounds plausible but is factually incorrect or completely made up. It might cite a study that doesn't exist, invent a statistic, or confidently describe something that never happened.

Why It Happens
LLMs predict the most likely text, not the most truthful text. They don't have a built-in fact checker. They generate what sounds right based on patterns in their training data — and sometimes what sounds right isn't actually right.

How to Reduce Hallucinations

•

RAG (Retrieval-Augmented Generation) — ground the model's responses in verified source documents

•

Fact-checking — verify claims against reliable sources before acting on them

•

Lower temperature — reduce randomness in the model's output to get more conservative responses

•

Ask for citations — prompt the model to provide sources so you can verify them

•

Human review — always have a person review AI-generated content before it's published or acted upon

The bottom line: This is why we can't blindly trust AI output — critical evaluation is essential. AI literacy isn't just about knowing how to use the tools. It's about knowing when to question them.

Multimodal AI

Multimodal AI refers to models that can process multiple types of input: text, images, audio, and video. Instead of being limited to reading and writing text, these models can see, hear, and work across different formats.

Examples

•

GPT-4o can see and analyze images you share with it

•

Claude can read and understand PDFs, images, and documents

•

Gemini can process video content and reason about what it sees

Specific model names will date quickly — capabilities ship every few months. The pattern is what matters: AI is moving from text-only to processing whatever you can show it.

Why does this matter? Because AI is moving beyond just text. The next generation of AI tools will understand the world more like we do — through multiple senses, not just words on a screen.

The Human Role — Why Learning Design Matters

AI can generate content, but it can't design how people learn. It can produce a lesson plan in seconds, but it doesn't understand what makes a concept click for a specific learner, or how to sequence ideas so they build on each other.

The biggest skill gap isn't syntax — it's strategy. Knowing which tool to use is easy. Knowing why, when, and how to use it effectively is what separates people who dabble from people who create real value.

Understanding AI fundamentals — what you just learned in this course! — is the first step. You now have the vocabulary and mental models to evaluate AI tools, ask better questions, and make informed decisions.

The real insight: The people who thrive won't be the ones who learn the most tools — they'll be the ones who learn how to learn with AI.

The AI landscape is evolving rapidly. Here are some of the most important trends to watch:

  • Smaller, faster models — Models are becoming more efficient, running on phones and laptops instead of requiring massive data centers. This makes AI more accessible and private.
  • AI agents becoming more autonomous — Agents are moving from simple task completion to handling complex, multi-step workflows with minimal human intervention.
  • AI in education — Personalized tutoring, adaptive learning paths, and AI-assisted course design are transforming how we teach and learn.
  • Regulation and safety — Governments worldwide are developing frameworks to ensure AI is used responsibly, with guardrails around bias, privacy, and transparency.
  • Open source growth — Open-source models like Llama and Mistral are making powerful AI available to everyone, not just big tech companies.

Key Takeaway

AI agents go beyond chat — they plan, use tools, and complete multi-step tasks. Hallucination is a fundamental limitation of how LLMs work, not a bug to be patched. The future belongs to people who understand both how AI works and how people learn.

Try This (Optional) — Course Capstone

Pick one task from your actual work that you'd like to delegate to AI. Sketch how you'd build it using techniques from at least three lessons in this course. Which model family would you reach for, and why? What would the system prompt say? Would you use RAG or fine-tuning — or both? Where in the workflow does a human review step belong? You don't need to build it. Describing it well is the harder skill.

Knowledge Check

What is the key difference between a chatbot and an AI agent?

Correct! The defining feature of an AI agent is its ability to plan, use tools, observe results, and adapt — completing complex tasks autonomously rather than just responding to individual messages.
Not quite. The key difference is that agents can plan, use tools, and complete multi-step tasks autonomously — they don't just respond to one message, they take action.

Your team is building a customer support chatbot using an LLM. Customers ask about your product's specs, which change every release. The bot keeps inventing features that don't exist. Which approach most directly reduces this kind of hallucination?

Correct! When facts change frequently and accuracy matters, RAG is the right tool — it pulls the current document into the prompt so the model has the actual spec to work from instead of guessing.
Not quite. The problem is that the model has no access to your current specs — it's filling in plausible-sounding text from its training data. RAG fixes this directly by retrieving the actual document at query time. Bigger models or longer prompts won't help if the facts simply aren't in the model.

You Did It!

You now understand the fundamentals of AI and machine learning. You know how neural networks learn, what makes language models work, how to prompt them effectively, and how techniques like RAG and fine-tuning customize AI for specific needs.

This isn't just knowledge — it's a foundation. Whether you're evaluating AI tools for your organization, designing learning experiences around AI, or just trying to cut through the hype, you now have the vocabulary and mental models to do it with confidence.

Back to Course Home Visit ashimasitaula.com

Was this course helpful?

Your feedback helps me improve the learning experience.

Previous: RAG & Fine-Tuning Back to Course Home