Architecture A vs Architecture B
Architecture A and Architecture B are two enforcement approaches for AI agent governance, defined by the AARM framework (Herman Errico, 2026). Architecture A places a protocol gateway between agents and their tools. Architecture B instruments governance inside the agent runtime. Each has structural tradeoffs in coverage, deployability, and visibility.
Why It Matters
Architecture A (protocol gateway) works today, with any agent platform, without requiring cooperation from the agent vendor or the downstream app. Every tool call passes through the gateway. Its blind spot: it cannot see what happens inside the agent.
Architecture B (agent platform hooks) sees everything the agent does, including intent, model context, and non-MCP side effects. But it requires the agent platform to expose governance hooks, and each platform's hooks are different.
Neither approach alone is sufficient. Architecture A provides universal coverage at the protocol boundary. Architecture B provides deep visibility inside the runtime. The AARM framework also defines Architecture C (kernel/OS-level) and Architecture D (tool vendor integration via XAA).