LoopArena Benchmarks the Model That Watches Your Coding Agent

Everyone runs coding agents in loops now. Almost nobody has asked, in a measurable way, who should run the loop. LoopArena, from Alibaba's AMAP team, is the first benchmark built around exactly that role: a Controller model reads structured progress updates after each iteration of a fixed Worker coding agent, then decides what to do next, or whether to stop. Paper is arXiv 2608.28281, and it topped Hugging Face daily papers with 82 upvotes; code and data at https://github.com/AMAP-ML/LoopArena.

Two numbers stand out. Best full-task success rate across all controllers tested: 24.69 percent. Supervising a loop turns out to be much harder than participating in one, and the leaderboard for it barely correlates with raw coding ability. Meanwhile good controllers cut inference cost by an average of 64.4 percent, mostly by knowing when to halt a loop that's going nowhere. Knowing when to stop is worth more than knowing what to type.

This slots straight into the harness thread we've been tracking all month: once the harness is the product, the scarce skill is loop judgment, exactly what practitioners have been doing by hand when they babysit long-running agents. A benchmark makes it trainable. Expect controller-tuned models within a quarter.