AI Agents, Clearly Explained

Level 1: Large Language Models (LLMs)
📌 LLMs, powering chatbots like ChatGPT and Gemini, excel at generating and editing text based on their training data.
⚙️ LLMs are passive, waiting for a human prompt to produce an output, and lack knowledge of proprietary or personal information.
🖼️ A simple LLM interaction involves a human input resulting in an LLM output (e.g., asking for an email draft).

Level 2: AI Workflows
🔗 AI workflows involve creating a predefined path or control logic for the LLM to follow, often using tools (like APIs or databases) sequentially.
🔄 If the workflow path is rigid (e.g., always searching a calendar), the LLM fails when asked a question requiring a different tool (like a weather API).
🛠️ Retrieval Augmented Generation (RAG) is identified as a specific type of AI workflow that enables models to look up information before answering.
🤖 In a workflow, the human remains the decision maker, manually iterating and rewriting prompts if the output is unsatisfactory.

Level 3: AI Agents
🧠 The critical shift to an AI agent occurs when the LLM replaces the human decision-maker within the process.
💡 An AI agent must autonomously reason (determine the most efficient approach) and act (use tools to execute tasks).
🔄 AI agents possess the ability to iterate autonomously, using techniques like the ReAct framework (Reason + Act) to critique and refine their own output until a goal is met.
👁️ An example shows an AI vision agent reasoning about what a "skier" looks like, acting by searching video footage, indexing clips, and returning results without manual pre-tagging by a human.

Key Points & Insights
➡️ LLMs are passive and limited to training data knowledge, unlike agents that can actively use tools.
➡️ AI Workflows are defined by human-programmed paths, whereas AI Agents feature the LLM as the decision maker.
➡️ The defining feature of an AI agent is its capacity to reason, act via tools, observe the result, and autonomously iterate to meet a specified goal.

📸 Video summarized with SummaryTube.com on Oct 28, 2025, 13:17 UTC