THE ANATOMY OF A PRODUCTION HARNESS (COMPONENT MAP)

CHECK YOUR UNDERSTANDING • Which component assembles the standing rules from global, project, and task sources? ◦ Answer: The instruction manager. • Why is a model adapter useful? ◦ Answer: It translates between the harness's common format and each provider's format, so you can change providers or models without rewriting the rest of the harness. • Why is the permission resolver written as ordinary code rather than left to the model? ◦ Answer: Because a model can be tricked or can make mistakes, so rules must be enforced by predictable code regardless of what the model says. • What three resources does a budget tracker commonly watch? ◦ Answer: Money, time, and tokens. • Describe the core idea of a harness in one sentence. ◦ Answer: It is a while-loop that calls the model, checks whether the requested action is allowed, runs it through the tool registry, records the result, and repeats until done or stopped. PART THREE: THE EXECUTION LAYER (RUNNING THE AGENT LOOP) In Part Two you learned that a harness is the supporting software wrapped around an AI model, and that researchers organize harness work into seven layers called Execution, Tooling, Context, Lifecycle, Observability, Verification, and Governance, abbreviated ETCLOVG. Part Three opens the first of those seven layers, the Execution layer. The Execution layer answers a simple but deep question: how does the software actually run an agent from the first moment to the last? Think of a restaurant kitchen. The chef (the model) has talent, but the kitchen (the harness) decides when orders come in, which station cooks what, what happens when a pan catches fire, and when the kitchen closes for the night. The Execution layer is the part of the kitchen that keeps the cooking going in an orderly way. In this Part you will study four things: the core loop that every agent runs (Chapter 9), how the harness remembers where it is and controls its pace (Chapter 10), how the harness decides which step comes next (Chapter 11), and the safe places where the agent's actions are carried out (Chapter 12).