Loop Engineering refers to the design of the "think-act" loop itself for AI agents: it encompasses context management, model selection, observation compression, and loop termination criteria.

The term "Loop Engineering" gained traction in the community in 2026. Kumar and Pandey provided a systematic definition in their paper TurnCost: Turns, Not Observations: Cache-Aware Loop Engineering for Coding Agents and conducted a quantitative analysis based on 600 public coding agent trajectories.

A core finding of Loop Engineering is that "the number of turns matters more than observations": the median length of a single observation is only 118–286 tokens, while large blocks of build logs, which are often compressed, only appear at the 99th percentile; 98.6% of input tokens hit the prefix cache, meaning that a "flat per-token pricing" model overestimates loop costs by an average of approximately 7 times. More critically, the number of turns explains 84% of the log-variance in trajectory cost, whereas the average observation size explains only 4%—the most expensive trajectories are typically those with the most turns, not those with the longest observations.

The significance of Loop Engineering is shifting the focus of agent efficiency from "how fast the model is" to "how well the loop is designed." Optimization should prioritize reducing "unnecessary turns" rather than solely focusing on the length of individual outputs; this also suggests that cache-aware cost models are closer to reflecting true costs.