From Trace Entropy to Coordination Control: Diagnosing and Intervening in Multi-Agent LLM Systems

Coordination interventions for multi-agent LLM systems are regime-dependent: identical guidance improves accuracy under greedy decoding but degrades it under stochastic decoding. We trace the reversal to over-regularization: exploration-promoting interventions add explicit entropy to policies already receiving implicit entropy from temperature, pushing effective diversity past the optimum. This leads to a decomposition of the intervention space into regime-dependent types (exploration, commitment) and regime-invariant types (consistency, disruption). We derive the Transfer Coefficient, a per-step information-gain diagnostic that detects coordination pathologies from the agent's own trace, and use it to build TRAC (Trace-based Regime-Aware Coordination). TRAC implements the regime-appropriate composition through two gated mechanisms: static self-trend constraints for failures the agent can self-diagnose, and a dynamic monitor for failures requiring external trajectory observation. Neither component alone reliably outperforms the baseline, but they address disjoint failure modes and their gated combination achieves 82.0% on GAIA versus 75.8% without intervention (${+}6.2\%$, McNemar $p{=}0.004$).