03 · Which rules bind · Rules # Rules are the operating language. A rule says what the seat may do, what it must hand on, and what still needs a person. It is not a personality, and it is not a longer prompt. [Try the exercise →](#try-it)[All three lessons →](/learn/) What it may do. What it must hand on. In this lesson [The idea](#the-idea)[What a rule binds](#what-binds)[Try one rule](#try-it)[The limits](#limits) ## The third root is the language of the work. A seat and a record still need a rule for how to act. Otherwise the agent improvises a procedure the team never agreed. In our team, the start check asks which rules are in reach before the task begins. The rule has to be where that check can read it. A personality describes a voice. A rule describes a boundary. ## Three lines are enough to begin. - May **What this seat may decide.** The bounded job, and no further. - Hand on **What belongs to another seat.** Name where it goes. - A person **What still needs a human.** Consequential calls stay with the person who answers. ## Try it with one rule you already follow. - **Write the rule** in the three lines: may, hand on, a person. - **Name the seat** it binds. The rule needs a seat that carries it. - **Put it where the agent reads rules** at the start of the work. - **Ask the agent to show the rule it is using.** A rule named in the answer is not proof it was followed. The seat check in the[first lesson](/learn/a-role-is-a-seat/#try-it) is where that reading starts. ## What this does not prove. A rule does not make an agent more accurate. A rule on a page the agent never opens does not bind the work. People still accept purpose, vision, strategy, objectives and key results. Instructions have to reach the agent, permissions have to hold, and the result still needs review. [Read the current limits →](/limits/) [The open research behind the lessons →](/research/)