digestVidSign in
summary

Jeff Su - AI Agents, Clearly Explained

Thumbnail of the video “Jeff Su - AI Agents, Clearly Explained”
Community upload. AI output has not been independently verified.
Author: Jeff Su
Categories: Technology Productivity
Tags: technology, productivity
Views: 2 · Last viewed: 2026-10-15 18:59 UTC

Short Summary

Main recommendation: Learn AI agents by building up from what you already know: LLMs (like ChatGPT), then AI workflows, then AI agents, where the key distinction is whether the LLM or a human makes the decisions. The video argues an AI agent is simply an AI workflow in which the human decision maker is replaced by an LLM that reasons, acts with tools, observes results, and iterates toward a goal. Its practical value is in understanding how this affects everyday AI use, without technical background.

Key Points

  • Level 1, LLMs: Chatbots such as ChatGPT, Google Gemini, and Claude are applications built on large language models. They are passive, responding only to prompts, and have limited knowledge of personal or proprietary data such as your calendar.
  • Level 2, AI workflows: A human predefines the path (the "control logic") the LLM must follow, such as searching Google Calendar before answering personal-event questions. They break when a question falls outside the predefined path, such as asking for the weather.
  • RAG (retrieval augmented generation): The video explains this as a process that lets AI models look things up before answering. It is simply a type of AI workflow, not a separate level.
  • Level 3, AI agents: The LLM becomes the decision maker. It reasons about the best approach, acts using tools, observes interim results, and autonomously iterates until the goal is met.
  • ReAct framework: The video describes this as the most common AI agent configuration, named for the requirement that agents both reason and act.

Details Worth Noting

  • Concrete instructions: The video does not give step-by-step setup instructions for the user. Its actionable advice is conceptual: to tell whether a system is a workflow or an agent, check who is the decision maker. If a human sets the path, it is a workflow even with hundreds of steps. If the LLM decides the reasoning and next actions, it is an agent. The speaker mentions a related video on building a prompts database in Notion that viewers might find useful.
  • The Make.com example: a workflow that compiles news links in Google Sheets, summarizes them with Perplexity, drafts LinkedIn and Instagram posts with Claude, and runs daily at 8 a.m. Because the human still sets the steps and rewrites prompts when the output is unsatisfactory, it is a workflow.
  • To turn that workflow into an agent, the LLM would need to choose tools (for example, picking Google Sheets over Microsoft Word for compiling links) and critique its own drafts against LinkedIn best practices in repeated cycles.
  • A demo of an AI vision agent searched video footage for a "skier" by first reasoning about what a skier looks like, then indexing matching clips, replacing manual review and tagging.

Bottom Line

The practical test is simple: if you set the steps, you are using a workflow; if the AI decides the steps, reasons about options, and loops until the result is good, it is an agent.