Anthropic's conflict tests show why multi-agent coordination needs mandate invariants, escalation rules, and independent acceptance checks.