What Really Happens When ChatGPT Does a Task for You (AI Agents Explained)

You type one sentence into ChatGPT's Work mode: "Find the cheapest flight to Goa next Friday, book it, and add it to my calendar." Then it searches, compares, opens websites, and fills in forms. Dozens of steps, on its own. So what's actually going on in there? In this video we switch between what you see on the screen and what's really happening underneath, on one simple map. You'll learn: • Why Chat mode answers but Work mode acts: same AI, plus a loop and tools • The agent loop: think → act → look → repeat • What a "tool call" is, and why the AI never clicks anything itself • How the agent uses its own virtual computer (screenshots in, clicks out) • Why long tasks drift: the notepad (context) has a size limit • Why small errors add up: 99% per step × 100 steps ≈ only 37% • Prompt injection: how a webpage can trick an agent • A real case from October 2026: Anthropic's test agents filled out government forms • The guardrails that keep agents safe: sandbox, minimal tools, human approval, logs ⏱ Chapters 0:00 Intro: one sentence, dozens of steps 0:32 Chat vs Work mode 1:29 The agent loop 2:59 Tool calls: what the AI really does 4:17 Its own virtual computer 5:20 Memory & why long tasks drift 6:11 When agents go wrong (prompt injection, real case) 7:53 Guardrails: sandbox, approvals, logs 8:57 Recap & when to use it Note: the ChatGPT screens are recreations for teaching, and the airline, flights and prices ("SkyFly", ₹6,240) are made up. Source for the October 2026 incident: Honolulu Star-Advertiser / AP, Oct 10, 2026. Grokable — tech you can actually understand. Subscribe for the next one. #AIagents #ChatGPT #AIexplained