CONTROL STRUCTURES AND DECISION ROUTING

DETERMINISTIC CONTROL VERSUS MODEL-DRIVEN CONTROL Now we come to one of the most important design choices in the whole discipline. Deterministic control means the harness's own fixed code decides what happens next. Deterministic means that given the same input, you always get the same result. A rule like "always run step A, then step B" is deterministic. Model-driven control means the model itself chooses what happens next, based on its reading of the situation. Neither is better in every case. Deterministic control is predictable, cheap, fast, and easy to test and audit. It is the right choice when the correct answer is known in advance and mistakes are costly, such as checking a user's permission or deciding whether to pause for approval. Its weakness is rigidity: it cannot handle situations its authors did not foresee. Model-driven control is flexible and can handle open-ended, messy tasks. It is the right choice when the right next step depends on understanding language or unexpected findings. Its weakness is variability: the model may choose differently each time, may make mistakes, and costs more because it requires a model call. An analogy is a hospital. The rule "wash hands before surgery" is hard-coded into policy, because it is safety critical and the same every time. But which diagnosis to pursue when a patient arrives with unusual symptoms is left to the doctor's judgment. A well-designed harness looks the same. It is a hybrid. The safety-critical backbone is deterministic, while the exploratory parts are model-driven inside limits. A helpful rule of thumb is this: if the decision protects money, safety, privacy, or other irreversible matters, lean deterministic or at least put a deterministic check around the model's choice. If the decision is about understanding messy content or choosing among many reasonable approaches, lean model-driven, but still keep step limits and guardrails. A guardrail is a rule outside the model that bounds what it may choose, for example "the model may pick any of these five tools, but never more than ten calls in total."