Manager-worker self-orchestration helps some models by 23 points and hurts others, with no way to tell in advance except measuring

arXiv 2608.26480·high signalXBlueskyLinkedInCopy link

arXiv 2608.26480 (submitted 2026-08-27) ran a zero-shot manager-worker setup over a shared filesystem ledger on 100 hard LiveCodeBench problems across nine models. Qwen3.8-27B gained 23.4 points, GPT-5.6-Luna 10.6, GPT-5.6-Terra 8.0, while Qwen3.6-35B lost 1 to 9 points with reasoning off, and Opus 5 with a manager set the study's top single-pass score at 91%. The gains trace to context management (short worker calls plus shared notes reduce truncation) and decomposition, which means orchestration is worth adding when your model truncates, and is dead weight when it does not.