WHY HARNESS ENGINEERING EMERGED AS ITS OWN FIELD

THE SHIFT FROM CHATBOTS TO AUTONOMOUS AGENTS To understand why a new discipline was needed, look at how the way people use language models has changed. The first widely used chat assistants, which arrived around 2022 and 2023, were essentially question-and-answer machines. You asked something, you got an answer, and the interaction ended. The cost of a mistake was small, because a human read every answer and decided what to do with it. The harness for such a chatbot could be thin: take the user's message, add some instructions, call the model, show the reply. Over the following two years, models became better at following instructions, at writing and fixing code, and at deciding when to use a tool. Builders therefore began to hand models longer jobs. Instead of asking, "How do I fix this bug?", people began asking an agent, "Fix this bug," and leaving it to read files, run tests, and make changes. By 2025 and 2026 it had become routine for agents to complete tasks lasting many minutes or hours, such as researching a topic across dozens of web sources, assembling a report, or refactoring a large body of code. Refactoring means reorganizing existing code to make it cleaner without changing what it does. This change in task length matters enormously. When a human reads each answer, small errors are caught immediately. When an agent takes two hundred steps with nobody watching, a small error in step ten can silently poison every later step. The agent may also run up large costs, touch files it should not, or loop forever. A thin harness cannot cope. Reliability, safety, and cost control become problems of the surrounding system, and they demand the sort of careful engineering that earlier chatbots never needed. Think of the difference between lending a friend your pen and lending a friend your car. The pen needs no rules. The car needs insurance, a license, fuel, a seat belt, and agreed limits. Autonomous agents are the car.