The sorcerer’s apprentice: understanding AI agents
About this app
An AI agent as Goethe’s sorcerer’s apprentice: the broom gets a task and runs the glowing loop plan, act, check, correct. Every round carries a bucket to the goal, mistakes it finds itself. Three everyday tasks (farm-shop website, club party, train journey), a leash of approvals for money, sending and deleting, and a stranger’s note in the bucket.
With explain mode: an AI tutor walks you through the app, points out what matters in the scene and answers your questions.
The idea
A chatbot answers, and you have to act yourself. An AI agent gets a goal and does the steps: it plans the next step, uses a tool such as browser, files, calendar, code, email or payment, checks the result and corrects itself. In programming this is called agentic coding: plan, build, test, repair. Because an agent really acts, it can also cause damage, for example send mail to the wrong people, delete files or book the wrong thing.
What the app shows
A broom, the apprentice from Goethe’s ballad, walks around a glowing circle from station to station. Every finished round pours a bucket into the basin, and on the right the result grows line by line. In the centre stands the master with the leash: confirm every step, ask only for money, sending and deleting, check at the end, or free hand. Without checks and with a free hand the water rises on delicate tasks.
Try it yourself
Choose a task: a website for a farm shop, a club party or a train trip. Watch the apprentice find an error while testing and repair it himself. Then set the leash to free hand and put a foreign note on a web page. Afterwards try the counter-spells: clear goal, approvals, test runs, log, small steps and undo.
Frequently asked questions
- What is an AI agent?
- An AI agent is a language model that works on a goal in steps by itself. It plans, uses tools such as a browser, files or a calendar, checks the result and corrects itself until the goal is reached.
- What is prompt injection?
- An instruction hidden in content that is read, for example on a web page or in an email. The agent may take it for part of its task. Approvals before money, sending and deleting are the most important protection.
- How do I use AI agents safely?
- With a clear goal, approvals before actions that cannot be undone, test runs, a look at the log, small steps and tools that allow undo. The human remains responsible.
- What does the app “The Sorcerer’s Apprentice” show?
- It shows an AI agent as a broom that plans, acts, checks and corrects in a circle and carries buckets into the goal. With a leash for approvals, three everyday tasks and a foreign note as a picture of prompt injection.
Subject: Understanding AI
More from Understanding AI
- The word oracle: how a language model writes
- Thinking muscle: sedan chair or training partner
- The beautiful facade: why AI answers sound so convincing
- Expectation and impact: overestimated in the short run, underestimated in the long run
- Magic box AI: one spark, six powers
- Spell workshop: talking to AI well
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps