When Does Coordination Induce Correlated Failure in Multi-Agent LLM Reasoning?

Multi-agent LLM systems increasingly rely on a meta-agent that issues a shared, explicit plan to coordinate otherwise independent executor agents, but whether such coordination improves or harms reasoning reliability is not well characterized. We present Controlled Deliberative Decentralized Execution, a minimal framework that isolates coordination strength as a single manipulable variable. We evaluate seven open-weight models across three reasoning domains. We introduce a two-statistic diagnostic framework—a correctness intraclass correlation ($\rho$) and an excess pairwise agreement statistic ($\kappa$)—designed to separate genuine correlated failure from benign narrowing of each executor's individual answer distribution. Strong coordination reduces accuracy relative to self-consistency in 19 of 21 (model, domain) cells, and this directional harm is unanimous at the model level: all 7 models are harmed on at least two of three domains. Critically, the underlying mechanism is not uniform: only three of seven models exhibit the correlated-cascade signature (rising $\kappa$) while the remaining four show comparable harm through a distinct, non-correlational route. This mechanistic split tracks neither model scale nor family. One robust exception survives across independent measurement corrections: Qwen2.5-3B on MATH shows a significant accuracy improvement under coordination, confirming that agreement's decoupling from correctness cuts in both directions. We argue that executor agreement is an unreliable proxy for reliability and that single-mechanism governance heuristics are miscalibrated for a meaningful fraction of deployed systems.