THE ANATOMY OF A PRODUCTION HARNESS (COMPONENT MAP)

THE PERMISSION RESOLVER The permission resolver is the component that decides what the agent is allowed to do. When the model asks to run a tool, the harness does not simply obey. First the permission resolver looks at the request and answers one of three questions. Is this allowed outright? Is this forbidden? Or is this allowed only if a human approves? The answer depends on rules that someone has written down, such as "reading any file in the project is fine," "writing files is fine inside the working folder only," and "pushing changes to the main branch needs a person's approval." Think of a building with a security desk. Some visitors may walk straight in, some are turned away, and some wait while the guard phones someone upstairs. The guard does not decide based on how polite the visitor sounds, but on the posted rules. That point is important for harnesses. The model is not trusted to police itself, because a model can be tricked or can simply make a mistake. The permission resolver is ordinary, predictable code that enforces the rules whatever the model says. Part Nine, Chapter 37, treats permissions and access control in depth.