What's this about?
The paper names a practice that has been forming without a name. Loop engineering is organising development work around coding agents: instead of writing each prompt by hand, you design a loop that monitors progress, assigns work, runs checks, and decides what the agent should do next. LoopArena benchmarks models as the *runtime controller* of that loop. 90 upvotes.
What is the difference between prompting and loop engineering?
Prompting is telling the model what to do. Loop engineering is building the thing that decides what to tell it, repeatedly, based on what happened last time. The model appears in two roles: as the worker doing the task, and as the controller judging whether the work is done, whether to retry, whether to escalate. Those are different skills, and a model can be excellent at one and poor at the other — which is exactly what a benchmark like this exists to reveal.
Why does this matter to you?
- It is the same distinction as automating versus augmenting from last week, moved one level up: doing the task versus running the process.
- The controller role is where reliability actually lives in production agent systems, and almost nothing measures it.
- "Loop engineering" is a useful term for what most people building on agents are actually doing.