Communication Boundary Failures in Multi-Agent Language Models

Multi-agent language-model systems can fail before any sentence is worded: at a ready-agent boundary, an agent must decide whether to act locally, yield, or communicate a recipient-targeted semantic record. We study communication-boundary failures, where this event-level decision is wrong even when the eventual text is fluent and schema-valid. Under one frozen improvement round with a fixed finite semantic interface, we define five operational failure modes–relevance–value confusion, generic-informativeness false negatives, proposal or shortlist loss, context aliasing, and wording leakage–and give minimal synthetic triggers plus trace diagnostics for each. The diagnostics separate recipient-side relevance from downstream value, proposal loss from value-estimation error, local-context aliasing from ordinary uncertainty, and parser equivalence from control equivalence. Existing controlled artifacts show that decision-focused relevance outperforms generic informativeness, semantic bundles recover value missed by top-1 proposals, and wording stress tests distinguish safe from leaky realization regimes. The repair is a screen–control–realize layer: prune only certified local negatives, choose among local action, yield, and retained sends by net value and cost, deliver canonical records to future control, and audit traces for missed acts, aliasing, and wording drift.