Why do AI agents need tools?

Ask an AI whether you need an umbrella today, and it might say probably not while it pours outside. Why can’t it look? On its own, it cannot. Switch the tools on, and it can.

▶ Try it yourself in Agent Tools: switch the tools on and off and watch the model move from guessing to looking it up.

What can a model do on its own?

A language model turns the text in front of it into more text, one token at a time, using patterns it learned in training. That data stops at a training cutoff, and the model has no live connection to the world. So on its own, it can only answer the umbrella question with a plausible-sounding guess.

How does a tool change the answer?

The app tells the model which tools exist. Each has a name, a description and its inputs, such as get_weather(city). When a tool would help, the model writes a structured request instead of an answer.

The app, not the model, runs the tool and pastes the result into the conversation. The model reads the result, then either answers or writes another request. The guess becomes a lookup.

Why is it called a loop?

Repeating this think, act, check loop until the job is done is what people usually mean by an agent. In research it is often called ReAct, for reasoning plus acting.

The loop has three beats. The model thinks, and decides a lookup would help. It acts, by writing a request. It checks, by reading what came back before it answers or asks for more.

Many assistants now have search or code switched on by default, so a lookup may already be happening behind an answer. The model is only one part of that system.

Arithmetic shows the same gap. Without tools, today’s models often get four-digit multiplication right, but they cannot check it. A tool can. The product of 1,234 and 5,678 is exactly 7,006,652.

Predict first, then try

Before you change anything, predict whether a model with its tools off will give a confident answer about the rain. Then check it in Agent Tools. Switch the tools off, then on, and watch the answer change from a guess to a lookup.

Explain to yourself which step makes the answer depend on evidence: the model’s own memory, or the result the tool returned. Then consider why real apps ask before they send, buy or delete anything.

What does the model leave out?

The model’s tag reads “illustrative tool calls.” The weather, forecast, calendar events, search results and confidence percentages are made up to show the flow. They are not from a real model run or a real forecast. The model also leaves out:

• Real tool calls. Real tool calls are usually JSON with named arguments. Apps can run several tools at once, tools can fail, and a model can misread a result.
• Confidence. Assistants do not show a calibrated confidence like the meter in the model. The meter stands for how much an answer rests on evidence rather than memory.
• Risk. An agent that can act can also act wrongly. Real apps ask before sending, buying or deleting.

The photos are AI-generated illustrations. The screen text, the calendar grid, the chat and every number are drawn by the model.

Sources: Wikipedia, Large language model; Wikipedia, Intelligent agent; Yao et al. (2022), ReAct: Synergizing Reasoning and Acting in Language Models; Schick et al. (2023), Toolformer: Language Models Can Teach Themselves to Use Tools.