LoopArena benchmarks the controller model separately from the coding agent it drives

arXiv·high signalXBlueskyLinkedInCopy link

arXiv 2608.28281 (28 August 2026) names 'Loop Engineering' as the practice of designing loops that monitor progress, assign work, run checks and decide the agent's next move, then points out that a single end-to-end outcome cannot tell you whether the loop or the agent was at fault. LoopArena evaluates how well one model guides a separate coding agent, isolating failure modes like trusting a stale progress note, skipping verification, misspending budget, or stopping before the task is safe to submit. If you run an outer loop over Claude Code or Codex, this gives you a way to test the loop's judgment without conflating it with the worker's ability.