New arXiv paper 'Stop Hand-Holding' defines loop engineering for AI coding agents

In mid-2026, a slogan reshaped how practitioners think about coding agents: stop hand-holding your agent, start designing the loop that prompts it. A new arXiv paper by Sandeco Macedo (cs.SE) formalizes this shift with the concept of a 'loop specification'—a bounded, reusable artifact with a trigger, goal, verification step, stopping rule, and memory. The loop is handed to agent harnesses like Claude Code or Codex, letting the agent pursue goals autonomously without step-by-step prompting. The paper distinguishes this external loop from ordinary programming loops and the internal perceive-act-observe cycle, positioning loop engineering as a new layer after prompt, context, and harness.

Macedo's contributions include a definition and scope, an anatomy and taxonomy of loop specifications with a five-level verification ladder, and a descriptive analysis of the Loop Library—a public corpus of 50 real loops coded by hand. The corpus reveals key maturity indicators: 70% of loops verify in the autonomous zone, and 74% name their terminal states, while automated triggering and durable memory remain underdeveloped. The paper also provides design principles and anti-patterns drawn from literature on self-correction, reward hacking, and model-as-judge fragility. It cautions against overselling the approach, noting limits like verification burden, comprehension debt, and risk of cognitive surrender. Loop engineering does not retire prompt engineering—both serve distinct use cases.