How AI agents use tools

What does one turn of the loop look like?

1. 1The program sends the goal, the history so far and the list of tools.
2. 2The model replies either with an answer or with a structured tool call.
3. 3The program validates the arguments and runs the tool.
4. 4The result — data or an error — is appended to the history.
5. 5The loop repeats with the new information.

The model is not executing anything. Every real effect happens in the program around it, which is where your limits, checks and logging belong.

Why do good tool descriptions matter so much?

The description is the only thing the model has when choosing between tools. 'Get data' is a coin flip; 'Return the last 30 days of orders for a customer, by email' is a decision. Ambiguous names and overlapping tools are the main cause of an agent doing something reasonable-looking and wrong.

• Name the tool after the intent, not the endpoint.
• Say what it returns, including what happens when nothing matches.
• Constrain the input schema tightly — enums instead of free text where possible.
• Avoid two tools that could both plausibly answer the same request.

How do agents fail?

Where should a human sit in the loop?

At the actions that are hard to undo. Reading is cheap to get wrong; sending an email, charging a card, deleting records and pushing code are not. A confirmation step at exactly those points buys most of the safety without making the agent useless.

Reversibility is a better rule than importance. Ask whether you could undo it in a minute, and gate what you could not.

How do you debug one?

1. 1Read the trace: every tool call, its arguments and its result.
2. 2Find the first step that was wrong, not the step where it became obvious.
3. 3Decide whether the model chose badly or the tool behaved badly.
4. 4Fix the description or the schema if it chose badly; fix the tool if it returned something misleading.
5. 5Re-run the same goal and confirm the earlier step now goes the other way.