Alignment Doesn't Compose
An ICLR 2026 paper proves that individually aligned agents amplify bias when composed into multi-agent systems. The architecture itself is the problem — not the agents. Worse, providing objective context accelerates polarization rather than reducing it.