AI Agents Explained

From chatting to doing

A chatbot answers a question and waits for the next one. An agent is given a goal and works toward it on its own for several steps. Ask it to research three laptops and put the specs in a spreadsheet, and it might search the web, open product pages, pull out the numbers, build the file and report back.

The word agent is old in AI. It simply means something that senses its environment and acts to reach a goal. In the late 1960s, researchers at SRI built Shakey, a wheeled robot that could plan a route and push boxes around a room. The project also produced the A star search algorithm still used in maps and games.

How modern agents work

Today's agents usually wrap a large language model in a loop. The model reads the goal, decides on a next step, and calls a tool, such as a web search, a calculator, a code runner or an app's interface. The tool's result is fed back in, and the model decides what to do next. The loop ends when the goal is met or the agent asks a person for help.

Agents may also keep notes or memory between steps and can split a big job into smaller tasks, sometimes handing parts to other agents. Coding agents that read a codebase, make changes and run tests are among the most widely used examples.

Where they stumble

Every step carries a small chance of error, and errors can stack up over a long task. An agent might misread a page, click the wrong button or confidently report success when something failed. Agents that read web pages or emails can also be fooled by hidden instructions planted in that content, a risk called prompt injection.

Giving an agent access to money, private files or the ability to send messages raises the stakes. Good systems limit permissions, log every action and pause for human approval before anything hard to undo.

Using agents sensibly

Start with tasks that are easy to check, such as gathering information, drafting documents or tidying files you have backed up. Give a clear goal, the boundaries of what it may touch, and what done looks like. Review the output as you would review work from a fast but new assistant. Agents can save real time, but responsibility for the result stays with you.

← GPUs: Why AI Runs on Graphics ChipsHow to Write Good AI Prompts →