raghav@dev:~

Guessing vs checking, the real difference with AI agents

Imagine asking a friend how to make chai.

They tell you: one cup of milk, one spoon of tea, one spoon of sugar, boil for five minutes. You follow it exactly. It’s too weak. Your friend never tasted it, so they’ll never know.

Now imagine your friend makes the chai themselves. They add the tea, taste it, and think “a bit stronger”. They add more, taste again, and stop when it’s right.

Same friend. Same recipe. The only difference is that the second time, they tasted as they went.

That’s the difference between asking an AI a normal question and using an AI agent.

What it looks like

A normal prompt is just a question:

How do I make one cup of chai, not too sweet?

You get a recipe back. The AI answers once and never finds out if it was right.

An agent gets three extra things.

A goal with rules:

Make one cup of chai, medium strong, lightly sweet.
Taste after every change. Add small amounts at a time.
Stop when it tastes right.

Tools, the things it is allowed to do:

add(ingredient, amount)
boil(minutes)
taste()

And a loop: do a step, check the result, decide the next step, repeat.

A run looks like this:

  1. Add milk and one spoon of tea, boil. Taste: too weak.
  2. Add half a spoon of tea, boil. Taste: strong enough, not sweet enough.
  3. Add half a spoon of sugar. Taste: just right. Stop.

The taste tool is the whole point. Take it away and the agent is back to guessing, just with a stove.

One warning

An agent is only as good as its tasting. If it can’t check its work properly, it will confidently serve you bad chai.

So when you hand work to an agent, don’t spend all your time on the instructions. Make sure it has a good way to taste.