Enterprise multi-agent architectures resolve to three orchestration patterns (hierarchical, peer-to-peer, broker-mediated) with materially different governance properties: hierarchical concentrates accountability at the orchestrator and is the easiest to audit but the most exposed to orchestrator-compromise; peer-to-peer distributes accountability and is the most resilient to single-agent failure but the hardest to audit; broker-mediated centralises the inter-agent communication path and is the most defensible against the cross-agent prompt-injection class. The choice of pattern is not a free architectural decision in 2026 because the EU AI Act's Article 9 risk-management requirements and the OWASP Agentic AI threat surface impose specific control obligations on each pattern. An enterprise should default to broker-mediated for new deployments above the high-risk threshold; hierarchical is acceptable for low-risk and contained deployments; peer-to-peer should be avoided in production agentic AI in 2026 unless the audit substrate is materially stronger than vendor-native baseline.

Multi-agent architecture playbook. 90-day review cadence. Watches: (1) the A2A (agent-to-agent) protocol's adoption trajectory through 2026 (claim AM-050 covers in detail), (2) Anthropic Managed Agents and OpenAI Operator's evolving multi-agent primitives, (3) emerging case-law and regulatory guidance specific to multi-agent failure attribution (currently underdeveloped; expect first major precedent in 2026-2027), (4) MCP (Model Context Protocol) adoption that affects how broker-mediated patterns get implemented.