WHAT A HARNESS IS

THE CONTROL PLANE The control plane is the layer of the harness that decides how agents run, communicate, and recover. The term comes from the world of computer networks, where engineers distinguish the data plane, which carries the actual traffic, from the control plane, which decides where the traffic should go. For agents, the idea is similar. The work itself, such as reading documents and writing summaries, is done by the model and its tools. The control plane decides which agent or model handles which job, in what order, with what permissions and budget, and what happens if something fails. Imagine an airport. Airplanes carry the passengers, which is the work. The control tower does not fly any plane, but it decides which plane takes off when, which runway to use, and what to do if a plane has a problem. The control tower is the control plane. In a simple single-agent harness, the control plane may be only a few dozen lines of code that loop, check limits, and handle errors. In a larger system with many agents, it becomes a serious piece of software. It performs routing, which means sending each task to the right worker. It performs scheduling, which means deciding when work runs. It performs recovery, which means retrying or rerouting when something crashes. And it enforces policy, which means applying rules about permissions and spending. Chapter 8 breaks the harness into the components that make up a typical control plane, and Part Ten explores the case of many cooperating agents. Everyday Analogy: The Brain and the Body A comparison to a person can make all of this feel natural. Think of the model as the brain. A brain is extraordinary at thinking, but a brain alone in a jar cannot do anything. To act, it needs a body. Nerves carry messages, muscles carry out movements, a skeleton gives structure, eyes and ears bring in information, and a reflex system protects the body from harm without waiting for conscious thought. The harness is the body and the nervous system. The tools are the hands and feet. The context builder is like the senses and attention, deciding what information reaches the brain. The memory store is like notes and a diary kept outside the head. The permission system is like the boundaries a responsible person respects, enforced by habit and by law. The budget tracker is like the sense of how tired one is or how much money is left in the wallet. The logger is like a journal. The analogy has limits, as all analogies do. A human brain has its own memory and continuous existence, while a model call starts fresh each time. But the analogy captures what matters most for beginners. Intelligence alone does not do work. Work needs a body that can act, remember, sense, and stay within safe limits. It also highlights an important point about responsibility. If a body harms someone because it was not controlled properly, we blame the lack of control, not only the thoughts. Similarly, when an AI system causes damage, the question is often not whether the model reasoned badly but whether the surrounding system should have prevented the action.