Communication Boundary Control for Safer Multi-Agent Language Agents

Multi-agent language-agent systems can fail before any message is worded: a ready agent may fail to warn, warn the wrong recipient, over-escalate weak evidence, expose control-relevant content through raw text, or trigger costly downstream behavior with a poorly scoped communication event. We study this pre-wording decision as communication boundary control: at a ready-agent boundary, the actor chooses a finite high-level act–local macro-action, yield, or recipient-targeted semantic send–and natural language is generated only after the semantic record has been selected. For one frozen improvement round with a fixed finite semantic interface, we derive the local control target and its diagnostics: predictive send-versus-yield usefulness is a relevance target, not a send rule; Stage II must choose by net downstream value and cost; a context-measurable KL-regularized actor must average hidden-boundary values before softmaxing; and parser-equivalent wording is safe only when it preserves control-relevant boundary dynamics. Controlled diagnostics test these safety-relevant failure modes: decision-focused usefulness separates from generic informativeness, semantic bundles recover value missed by top-1 proposals, and wording regimes expose leakage from surface text. The result is a finite, auditable control layer for deciding whether a communication event should exist, who should receive it, and what control-relevant content future agents may condition on before wording begins.