Reasoning Between Actions: Prompted and Native

Should an agent state what it has learned before it chooses its next tool?

The agent loop does not require this. The model may request the next tool as soon as it reads a result, with no text in between. A short written comparison at the right moment can stop a wrong action, but it also adds tokens to every later step, and the designs that produce it differ in cost, visibility, and control.

The Running Case

A customer writes to Desk:

You charged me twice for my kettle. Please refund one charge.