Tool use is underrated, and everyone's overrating it

Tool use is the single biggest multiplier an LLM app can get. It's also the feature most likely to make your product worse if you deploy it wrong.

Why it's underrated

Most apps we see still treat the model as a text-in, text-out box. They don't give it the ability to look something up, do a calculation, or call the real function that knows the right answer. So the model makes things up — because that's what we asked it to do.

Three well-chosen tools beats a better model on almost any real task.

Why it's overrated

The teams that go too far the other way give the model a menu of twenty tools and let it pick. Every interaction becomes a non-deterministic search across the whole menu, and the wrong tool gets called half the time. The product becomes unreliable in a way that's almost worse than no tools at all.

The middle

Fewer tools, called with very specific guidance. Every tool should have:

• A name that states what it's for, not what it does.
• Parameter names that match the user's vocabulary, not the API's.
• A description that tells the model when not to call it.

If you can't write that description clearly, the tool doesn't belong in the menu yet.