Understanding AI Agents in 15 Minutes

Everyone is selling "AI agents." Very few people can say precisely what one is.

This book can, and it does so in a fifteen-minute read. It starts with an analogy (a brilliant consultant who may never leave the meeting room, until you hand her a telephone and a short list of people she may call), turns that analogy into an exact definition, and then shows the definition working: a complete agent in about sixty lines of Python, three of those agents arranged into a workflow, and the unedited output of a real run against Claude, six model calls from question to verified answer.

Along the way it draws the one distinction that matters most in practice: who holds the steering wheel. A workflow keeps the order of steps in your code; an agent hands it to the model. Real systems mix the two, and once you can see which is which you can read any framework, evaluate any vendor, and predict what a system will cost and how it will fail.

The book is honest about limits. A chapter on when not to use an agent covers cost, unpredictability, side effects and prompt injection. Another maps the no-code routes (chat-app agents, visual builders, open-source studios) back onto the same four ideas, so readers who will never write Python still leave knowing exactly what the products they use are doing. It closes with a one-page checklist for building a first agent this week, a glossary, and the full program, which is also available under the MIT licence at github.com/juliusbsimon/understanding-ai-agents.

Written for software engineers meeting agent architecture for the first time, architects and technical managers separating agents from ordinary automation, product people who need to challenge an "agentic" plan, and teams who want a shared vocabulary before an AI project starts. It is the guide to hand someone before the agent meeting.