An AI that takes a goal and works through many steps on its own, using tools.
A chatbot gives one answer per question. An AI agent takes a goal, breaks it into steps, uses tools such as search, files or code, looks at what happened, and decides what to do next, again and again.
The basic loop of thinking, acting and observing was laid out in the 2022 ReAct paper. In 2023, AutoGPT made "just give it a goal" hugely popular, but also showed agents getting stuck in loops and running up costs.
The more steps, the more one early mistake can snowball. For things you can't undo, like paying or sending messages, keep a human check at the end.
Think of it like this
One friend tells you how to cook instant noodles; another boils the water, adds the noodles and tastes the soup. A chatbot is the first friend; an agent is the second. But if you hand over the kitchen, you still check the stove is off.
For example
Ask it to "find and summarize sources for my science presentation" and it will search, read and summarize, step by step.
Common mix-ups
Mix-upAn AI agent is just a smart chatbot, so you can leave everything to it.
ActuallyIn 2024 Anthropic advised distinguishing "workflows", which follow fixed steps, from agents that choose their own next move, and starting with the simpler one. The same goes for you: hand over small jobs first and watch where it slips up.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- How do you split what you hand to an agent from what you must check yourself at the end?
- Activity
- Imagine an agent doing "find and organise material for next week's science presentation". Write the think → act → look-at-the-result loop as step cards (search, read, summarise…). Mark each card "can it be undone if wrong?", and slip a "person checks" card in front of every card that can't. Check the ReAct record (E0169) to see when the basic loop was set out.
Records about this
Source-checked moments from AI history.
Distinguished workflows from agents and recommended starting with direct API calls rather than frameworks; a summing-up of 2 years of hype
Created the story of an AI that just gets on with it once you give it a goal, while also exposing the public to infinite loops, hallucinations and runaway costs
The standard form of the agent loop that combines reasoning techniques with tool use. Cross-referenced with axis 3