AI Agents, Clearly Explained

This video is designed for individuals without a technical background who regularly use AI tools and want to understand AI agents. It simplifies intimidating terms like RAG and ReAct by building on concepts like ChatGPT and progressing through AI workflows to AI agents, using real-life examples.

Popular AI chatbots like ChatGPT and Google Gemini are built on LLMs, which excel at generating and editing text based on their training data. However, LLMs have limitations: they lack proprietary information (like personal calendar data) and are passive, only responding to prompts. This highlights their limited knowledge and reactive nature.

AI workflows involve humans defining a predefined path for an LLM to follow, such as searching a Google Calendar for event information. If the query falls outside this predefined path (e.g., asking about the weather), the LLM will fail because it cannot deviate from its programmed control logic. Even with multiple steps and API integrations, a workflow remains a workflow as long as a human is the decision-maker. Retrieval Augmented Generation (RAG) is explained as a type of AI workflow where the model looks up external information before answering.

An example of a real-world AI workflow using make.com is demonstrated. It involves compiling news article links, summarizing them with Perplexity, drafting social media posts with Claude, and scheduling the posts. This workflow follows a human-defined sequence of steps. The key takeaway is that a human manually iterates and refines the process, such as rewriting prompts for better output.

An AI agent differs from a workflow in that the LLM replaces the human decision-maker. The AI agent must autonomously reason about the best approach to achieve a goal and then act by using tools to execute those steps. This includes iterative refinement, where the AI agent can critique and improve its own output, similar to the ReAct framework (Reason and Act). Instead of a human manually rewriting prompts, an AI agent would use another LLM to critique and refine its output based on best practices until the goal is met.

Andrew Ng's demo website illustrates an AI vision agent. When a user searches for 'skier,' the agent first reasons what a skier looks like and then acts by searching video footage, indexing clips, and returning relevant results. This process is entirely autonomous, without human pre-tagging or manual review, highlighting the agent's ability to reason, act, and iterate independently to achieve a goal.

A visual summary of the three levels: LLMs take input and provide output. AI workflows take input and follow a human-defined path, potentially using external tools. AI agents receive a goal, and the LLM itself performs reasoning, takes action, observes interim results, and iterates autonomously to achieve the final goal. The crucial difference is that in AI agents, the LLM is the decision-maker in the workflow.