You may have heard the word "agent" a lot lately. It can sound mysterious, but the idea builds directly on what you learned about tools.
A model in a loop
An agent is a model that uses tools in a loop to work toward a goal. Instead of one request and one reply, it goes around and around:
- Look at the goal and what has happened so far.
- Pick a next step, often a tool call.
- Your code runs the tool and returns the result.
- Check the result, then go back to step 1.
The loop ends when the model decides the goal is done, or when your code stops it, for example after a set number of steps.
Say the goal is "find three nearby plumbers with good reviews and list their phone numbers." An agent might search the web, open a few pages, note what it finds, search again for missing details, and then write the list. No single step is clever. The value comes from stringing the steps together without you steering each one.
What agents are good at
Agents can save time on tasks that have several steps and a result you can check. Gathering information from many places, sorting through a pile of documents, or trying a fix and testing whether it worked are good examples. This is the same skill as breaking a task into parts, except the model does the breaking and the doing.
Where they go wrong
Agents are useful, but they are not magic, and they fail in some common ways:
- Going off track. The model can misread the goal or chase a side detail, spending many steps on the wrong thing.
- Compounding errors. Each step builds on the last. One early mistake, like a wrong date, can carry through everything after it.
- Cost. Every step is another request, and the history grows each time. A long run can use far more tokens than you expect.
How reliable agents are depends a lot on the task, the model, and the tools. It is still an area where people are learning what works.
Keep people in charge
The safest agents have limits built in by the people who make them. Cap the number of steps and the amount you're willing to spend. Give the agent only the tools it needs. Most of all, make it stop and ask before anything risky or hard to undo, like sending messages, spending money, or deleting things. The agent can propose. A person decides.
Try it
Pick a small errand you did online recently, like booking a table. Write out the steps you took. Then mark any step where you'd want an agent to stop and ask you first.