FRONT COVER MATTER

A PLAIN-LANGUAGE PREVIEW OF AI HARNESS ENGINEERING Before the detailed chapters begin, here is the whole idea in one short story. An LLM is a computer program that has learned patterns in enormous amounts of written text, and that can read some text and produce more text in response. You can think of an LLM as an extremely well-read engine. On its own, though, an engine does not go anywhere. An engine sitting on a workshop floor cannot carry you to the store. It needs a frame, wheels, a steering wheel, brakes, a fuel gauge, and a driver's seat. In the same way, an LLM on its own only turns text into text, and it remembers nothing from one run to the next. If you want it to behave like an agent, you must build things around it. An agent is a software system that can take actions on its own toward a goal, such as researching a topic, fixing a bug in a program, or answering a customer. The software you build around the model is called the harness. Harness Engineering is the discipline, meaning the organized professional practice, of designing the environment, the rules, the tools, the safety checks, and the feedback loops that surround an AI model so that it works reliably in the real world. This book uses a helpful three-layer picture throughout. First, the Model is the engine. It is the LLM that does the raw reasoning. Second, the Scaffold is the blueprint. It includes the system prompts, which are the standing instructions given to the model, the design of its memory, and the logic of the loop that tells the agent how to think through a task. Third, the Harness is the hands and feet. It is the secure software that actually runs tools, stores memory, manages what the model can see, recovers from errors, and keeps the agent inside safe limits. Researchers have also proposed a useful organizing framework that splits the harness into seven layers. It is called ETCLOVG, which stands for Execution, Tooling, Context, Lifecycle, Observability, Verification, and Governance. Do not worry about memorizing these words yet. Each one gets its own detailed treatment, and Chapter 7 introduces all seven together. This book uses the seven layers as its backbone, and it adds extra Parts for foundations at the start and for building, deploying, and the professional path at the end. A note on one important standard. The Model Context Protocol (MCP) is an open standard that describes a common way for AI applications to connect to outside tools and data sources. You will study it closely in later Parts. As reported by industry trackers in late 2026, the MCP ecosystem has grown to thousands of public servers (reports vary, and the exact count is uncertain), and the software development kits (SDKs), which are ready-made code packages that help programmers use the standard, are reported to be downloaded tens of millions of times per month. Anthropic, the company that created MCP, transferred its governance in December 2025 to a vendor-neutral foundation, reported as the Agentic AI Foundation under the Linux Foundation. The word vendor-neutral means that no single company controls the standard, which is good for builders who do not want to be locked in. These developments strengthen the chapters that teach MCP as an open, multi-vendor standard rather than change them.