Prompt-Only Meeting Delegates Stay Silent on 51.4% of Their Turns; an Architecture Cuts That to 2.5%

arXiv 2609.03923·medium signalXBlueskyLinkedInCopy link

LLM agents delegated to attend a meeting fail at knowing when to speak, with no structured way to track stances, coverage and floor, and prompt-only delegates stay silent through 51.4% of the absent participant's talking opportunities on the AMI corpus. CAPA splits the job across a Perceiver updating meeting state, a Predictor forecasting the conversation, a Controller deciding whether to speak and which proposition to surface, a Generator phrasing it in the participant's style, and two judges plus a Recalibrator scoring forecast and action against the next observed turn. On 137 AMI meetings it drops silence to 2.5%, doubles credited recovery from 26.1 to 52.2, and holds hallucination at 0.6%, with judges aligning to human annotation at Cohen's kappa 0.71.