BoardroomAI keeps the human in a multi-agent deliberation by compiling interventions into a typed decision graph

arXiv·medium signalXBlueskyLinkedInCopy link

BoardroomAI (arXiv 2608.13046, 2026-08-13) rejects the transcript pattern where a human states a problem, agents deliberate internally, and a final answer comes back. It models deliberation as a typed decision graph of evidence, assumptions, constraints, claims, objections, alternatives, risks and decisions with semantic dependencies and specialist ownership; an intervention compiler turns confirmed human actions — challenging an assumption, changing a constraint, adding evidence — into explicit graph updates, and dependency-aware propagation reactivates only affected specialists. Across 600 generated decision-DAG interventions, propagation matched exhaustive impact computation while inspecting 14.59% of nodes, and a 12-case pilot recomputed 62.11% of canonical nodes while preserving every gold-unaffected node.