Loop Freedom

Loop freedom is how much of the route the model chooses. Every agent loop sits on one axis: at one end the model gets a broad goal and finds its own path; at the other it runs a path a human designed first, with four parts: defined steps, a check after each one, a hand-off point for when it gets stuck, and a stop condition. Both ends run autonomously; the difference is who chose the route. Practitioners call the ends open and closed loops; the axis matters more than the label.

The trade is the same one as workflow versus agent, one level up. A free loop finds things you did not specify, and pays in token burn, drift from the goal, and weak control. A designed path is cheaper, more reliable, and testable, because every step you fixed is a step you can assert on.

The design rule: fix the definition of done before you free the route. For long-running work the baseline is a durable record rather than route freedom: what must be true, written down, with one test per statement, and the outcome of every attempt written back into it. Given that record, whether the loop continues on its own is a convenience, because each pass works out what is unblocked from what is currently true instead of following a remembered plan. The steps in between get designed only when the task repeats and reliability beats discovery, which is workflow territory. How much route freedom you can afford depends on how good your checks are, because a free loop converges on whatever its verifier fails to notice.