Tools Are Words

Overview

We know that the tool set is part of the conversation, and that the conversation is made of words. So what is a tool, to an LLM? Everything it will ever know about a tool arrives as text on the wall — a name, a description, parameters — and a name is a special kind of text: a single word that unpacks into knowledge the model already owns, at no cost to the conversation. In this lesson we weigh words like an engineer: what a well-chosen name gives the conversation for free, what a description must pay for every single turn — and what happens to the trajectory when a name lies.

Words decode into rich meaning

Say the word apple. That one small symbol decodes into something much richer: what an apple looks like, feels like, smells like, tastes like; where it grows; what it costs; what you can do with it. A tiny word, and a rich representation unfolds — in your brain, and in an LLM's weights — learned long before this conversation began.

A word you can't decode

Now the other side of the experience — and in this room, I'm the demonstration. Here is a Norwegian word:

Everyone who speaks Norwegian decodes that instantly. I don't speak Norwegian — in my head, eple decodes into nothing. No color, no smell, no weight. For me to understand it, someone has to pay words: "Jules, eple is Norwegian for apple." One line — cheap, because the concept already existed in my head and only the label was missing.

But now try a concept that doesn't exist in my language:

The point is simple: words you know carry rich meaning with no description. Words you don't know cost a description — and even a good description delivers less than native understanding.

This is precisely the situation between you and an LLM every time you hand it a tool. A name it recognizes is eple: the concept is already in the weights, and the label alone unlocks it. A tool it has never seen is utepils: you must pay a description. A familiar name is a pointer into the model's trained knowledge — the only context you never pay for. A description, by contrast, is a definition written into the conversation — and, as we know from the economics of the loop, one that is re-sent on every single iteration.

Compare two tool names

The same operation, offered to an agent two ways. First, a name the model doesn't know:

Now a name the model already understands:

Both tools do exactly the same thing. But names the LLM already understands are more token-efficient: they need no description. And because tool definitions are re-sent on every loop iteration, the saving repeats every turn — a five-step agent pays Q14875's description five times, some 225 tokens spent re-teaching what fetch_page(url) says for free. The best description is a name that needs none; pay tokens only for what the name can't say.

Known words carry their uses

There's more in a word than appearance. You also know what can be done with an apple — eat it, slice it, throw it, wedge something with it — and nobody ever handed you a manual titled "Actions Available for Apples." Watch an agent get the uses for free:

A tool the model has never met carries none of this: every use, rule, and limit must be paid for in the description.

Computers are full of unknown words

Here's the friction that makes this lesson matter. Using a tool well requires understanding it: what it can do, what it can't, how it combines with other tools. The more well-understood the tool, the less we have to write — and the less can be misread.

But computing has an inconvenient habit: we build one-off tools. Every internal API, every lab instrument script, every bespoke database schema is a novel word that exists nowhere in the model's training. That puts every tool somewhere on a spectrum:

Well-known tool names — where the tool behaves exactly as the name suggests — are cheaper. So the practical craft: name the novel thing after the nearest familiar concept that doesn't lie, and spend description tokens only on what is genuinely new.

The lying name

But the pointer cuts both ways. The model uses the word, not your code. As far as the trajectory is concerned, the tool is its name and description — the implementation is invisible until results come back, and sometimes not even then:

Rename the tool archive_file(path) and the same agent tells the truth — with zero changes to the code. Names are promises the conversation believes. A good name is free capability; a lying name is an agent that is wrong with confidence.