Harness per agenti: dal prompt al loop di strumenti

Abstract (EN)

An agent is not a bigger prompt: it is a loop. This article describes the agent harness, the software that wraps a model in an observe-decide-act cycle, lets it call tools, feeds results back, and decides when to stop. We separate the model's role (choosing the next action) from the harness's role (executing tools, managing context, enforcing limits) and catalogue the common failure modes: context overflow, tool errors mistaken for reasoning, and loops that never terminate. The aim is a clear mental model that makes agent behaviour debuggable, by locating each fault in either the model or the surrounding system.