Loop Engineering

The practice of organizing product development as cycles of execution, evaluation, and adjustment. In the fastest cycle, an agent implements a specification, runs tests, and fixes what failed. At longer intervals, the developer reviews the product and reframes decisions; feedback from users and the operating environment determines which hypotheses are still worth building.

The value comes from closing each cycle with a verifiable signal, not from keeping an agent working for longer. Specifications, evals, observability, stopping criteria, and rollback boundaries define the loop. Without them, automation merely speeds up changes whose quality and direction remain unknown.