Skip to main content
Back to Course 101
9/12
Unit 9 of 12
Unit 09 — Phase 03

AI That Does Things

Watch It Act. Then Build.

Listen to this unit

Read-aloud narrates the unit in a natural voice, paragraph by paragraph. It needs an account.

In short

AI agents go beyond conversation — they plan, use tools, execute tasks, and check their own work. The progression is: chatbot → assistant → agent. We're in the transition between assistant and agent.

Socratic Mode

The Hook

Everything you've used AI for so far follows the same pattern: you type something, it types something back. You ask, it answers. It's a conversation.

But what if AI could actually do things? Not just talk about sending an email — actually send it. Not just suggest a plan — actually execute the steps. Not just write code — actually run it, test it, fix the bugs, and try again.

That's what AI agents do. They go beyond a . And the shift from "AI that talks" to "AI that acts" is the biggest jump in capability since AI learned to use language in the first place.

Video — The Escalation of AutonomyWatch on YouTube

The Core Concept

An is an AI system that can break a goal into steps, use tools to complete those steps, check its own work, and loop until the job is done — with or without human supervision.

The key word is "tools." A chatbot can only generate text. An agent can search the web, run code, read files, interact with apps, browse websites, and call other s. It's the difference between someone who can tell you how to cook dinner and someone who can actually walk into the kitchen and cook it.

Explained your way: the AI rewrites this idea around something you already know.

Lilian Weng at OpenAI wrote what many consider the definitive blog post on AI agents, breaking them down into three components: Planning (how do I break this goal into steps?), Memory (what have I already tried, and what did I learn?), and (what external capabilities can I call on?).

Here's a concrete example. You tell an AI agent: "Find the cheapest flight from Amman to London in June, compare three options, and send me a summary by email." A chatbot would generate a paragraph about how to search for flights. An agent would actually go to flight comparison websites, search for dates, compare prices, format a summary, and send you an email — all autonomously.

In October 2024, Anthropic announced that Claude could use computers — literally. Claude became the first frontier AI model to operate a computer in public beta: it looks at the screen, moves the cursor, clicks buttons, types text, and navigates software interfaces.

In demonstrations, Claude filled out forms, navigated multi-step web workflows, and used desktop applications — all by "seeing" screenshots and deciding where to click next. The system isn't perfect: it's slow compared to a human, sometimes clicks the wrong button, and can get stuck in loops. But it crossed a fundamental threshold: AI that can use any software a human can use, without needing a custom API.

This matters because most business software doesn't have APIs. The only interface is the screen. An AI that can use the screen can, in principle, automate anything a human does on a computer.

OpenAI followed with their own agent infrastructure: the Responses API and Agents SDK, along with built-in tools for web browsing, file analysis, and code execution. Google, meanwhile, integrated agent capabilities into Gemini.

Then there's Devin, built by Cognition Labs, described as the "first fully autonomous AI software engineer." Devin can set up coding environments, search documentation, write code, run tests, debug errors, and submit pull requests — all on its own. Whether Devin lives up to the marketing is debated (some demos were found to be misleading), but the direction is clear.

Agents aren't chatbots with better language skills. The difference is structural. A chatbot generates text in response to input. An agent plans a sequence of actions, executes them using external tools, observes the results, and adapts its plan based on what happened.

This means agent failures are fundamentally different from chatbot failures. When a chatbot is wrong, you get bad text. When an agent is wrong, it can take bad actions — sending the wrong email, deleting the wrong file, purchasing the wrong item. The stakes are categorically higher, which is why design is so important.

The critical concept to hold onto is human-in-the-loop. Right now, the best agent workflows keep a human checkpoint at key moments — approve this before sending, review this before publishing, confirm before spending money. Fully autonomous agents make mistakes, and those mistakes compound: an error in step 2 can cascade through steps 3, 4, and 5 before anyone notices.

The progression goes like this:

Chatbot → answers questions when asked Assistant → answers questions + uses tools when directed Agent → receives a goal + plans steps + uses tools + executes + checks its own work

We're currently somewhere between assistant and agent. The trajectory is clear. What matters now is understanding the capabilities and the limits — because the limits are where the humans still matter most.

Knowledge check
What fundamentally separates an AI agent from a chatbot?
Knowledge checks save to your account.

Live Demo

Step 1: Open Claude (claude.ai) or ChatGPT. Give it a multi-step task:

Prompt
Research the top 3 AI companies by revenue in 2025, create a comparison table with their revenue, number of employees, and flagship product, and write a one-paragraph analysis of who's winning and why.

Watch how it breaks the task into steps, searches for information, and synthesizes.

Step 2: Now try a tool-use task. Ask Claude to analyze a piece of text:

Prompt
Read the following article [paste a short article] and create a bullet-point summary, identify the three strongest claims, and rate each claim's credibility from 1-5 with an explanation.

Notice: it's not just answering — it's processing, analyzing, and structuring.

Step 3: If you have access to ChatGPT's agent mode, try:

Prompt
Find a recipe for a chocolate cake that takes under 45 minutes, convert the measurements to metric, and format it as a printable recipe card.

Watch it browse, convert, and format.

Step 4: Try deliberately giving it a task where it should ask for clarification instead of guessing:

Prompt
Book the thing for next Thursday.

Does it ask for details, or does it guess?

Chatbot (text only)
You ask: "How do I find cheap flights to London?" The AI generates a helpful paragraph about flight comparison websites, tips for finding deals, and when to book. You read it, then go do all the searching yourself. The AI talked about the task but didn't do any of it.
Agent (text + action)
You ask: "Find the cheapest flight from Amman to London in June and email me a comparison." The AI searches flight websites, compares prices across dates, creates a formatted comparison table, and sends it to your email. The AI didn't just discuss the task — it completed it.

Why This Matters

The shift from chatbots to agents changes what's possible for individuals. Tasks that used to require a team — research, data analysis, multi-step project execution — can now be initiated by one person and executed by an AI agent.

But it also changes the stakes. A chatbot that gives bad advice is annoying. An agent that takes bad action — sends the wrong email, deletes the wrong file, publishes the wrong content — creates real damage. The evaluation skills you learned in Unit 08 become even more critical when AI is acting, not just talking.

The students who understand agents early will have a genuine advantage, because they'll know both what to delegate and what to keep hands-on. That judgment — when to let the agent run and when to stay in the loop — is the new management skill.

Knowledge check
Why is "human-in-the-loop" especially important for AI agents (compared to chatbots)?
Knowledge checks save to your account.

The Challenge

Agent Workflow Designer

35 minutesHands-on

Design and test a real multi-step workflow using AI:

  1. Identify a real task — Pick a multi-step task in your life (planning an event, researching a purchase, organizing notes for a class, creating a study schedule).
  2. Break it into 5-8 steps — List each discrete step the task requires.
  3. Assign each step — For each step, decide: should a human do this, should AI do this, or should AI do this with human review?
  4. Execute — Use an AI tool to actually execute the steps you assigned to AI. Document what worked and what didn't.
  5. Post-mortem — Write a brief review: Where did the AI save time? Where did it need correction? Where would you change the human/AI split?
Success criteria: You designed a real workflow with clear human-AI task division, executed it, and can articulate where AI was strong, where it needed oversight, and where a human was irreplaceable.
Submitting your work needs an account.

Key Takeaways

  1. 1AI agents go beyond conversation — they plan, use tools, execute tasks, and check their own work.
  2. 2The progression is: chatbot → assistant → agent. We're in the transition between assistant and agent.
  3. 3"Human-in-the-loop" is essential: agents make mistakes that compound across steps. Always build in checkpoints.
  4. 4The new skill isn't just using AI — it's knowing what to delegate and what to keep human.

The Rabbit Hole

Type: Article Title: A Practical Guide to Building Agents — OpenAI URL: https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf Description: The clearest overview of agent architecture, design patterns, and guardrails, from the company pushing hardest on building them. Free PDF.

Explore Further

TypeTitleURLDescription
ArticleLilian Weng, "LLM Powered Autonomous Agents" (foundational post)lilianweng.github.io/posts/2023-06-23-ag…The definitive technical overview of agent architecture
ArticleAnthropic, "Introducing Computer Use" (Oct 2024)anthropic.com/news/3-5-models-and…Claude becomes the first frontier model to operate a computer
ArticleOpenAI, "New Tools for Building Agents" (March 2025)openai.com/index/new-tools-for…OpenAI's agent infrastructure: Responses API and Agents SDK
GuideOpenAI, "A Practical Guide to Building Agents" (free PDF)cdn.openai.com/business-guides-and…Comprehensive guide to agent design patterns and guardrails
ReferenceWikipedia, "Devin AI"en.wikipedia.org/wiki/Devin_AIThe "first autonomous AI software engineer" — capabilities and controversy
ToolClaudeclaude.aiAnthropic's AI assistant with computer use capabilities
ToolChatGPT (Agent Mode)chatgpt.comOpenAI's AI assistant with browsing and code execution
ToolOpenAI Playground (tool use / function calling)platform.openai.com/playgroundTest agent capabilities with function calling
BookEthan Mollick, Co-Intelligence (2024)Framework for understanding AI as a collaborator with agency

Last updated: May 21, 2026. Video embeds and links were refreshed on this date; the agent concepts still hold.

Track your progress

Marking a unit complete, the Prove It check and your place in the course all need an account. The reading stays free.