The author observes a growing trend of building systems on top of coding agents, creating loops that extend the agent's capabilities beyond a single session. These loops involve task queues, machine execution, and a harness that decides whether to continue the task by modifying context or sending it to another machine. While coding agents have internal loops for tasks, a new external "harness-level" loop is emerging and dominating discourse. The author personally struggles with this hands-off approach for code they care about, finding the generated code too defensive, complex, and lacking strong invariants. Current models tend to add excessive local defenses rather than preventing undesirable states, a problem that seems to be worsening. Loops amplify these tendencies, making systems less understandable while appearing more robust. However, loops excel in domains like code porting, performance exploration, and security scanning, where code transformation or short-lived artifacts are the goal. In these cases, the generated code's longevity or verifiability is less critical than the process of exploration or transformation. The author contrasts this with writing lasting code, which they liken to evolving from software as a deterministic machine to software as an organism. This shift implies a move away from deep human comprehension towards monitoring, stabilizing, and treating systems almost like biological entities. Opting out of this machine-driven future is becoming difficult, especially in security, where attackers and researchers already use loops, forcing defenders to adapt. Competitive pressures and the speed of some teams leveraging orchestration will also drive adoption. The most concerning aspect is the potential cognitive and practical dependency on these looping systems, raising questions about future access, cost, and the ability to understand code without machine assistance. Ultimately, codebases may become reliant on machine participation for their maintenance.