How to implement agent routing and autonomy: intent dispatch at ingress, orchestration in the loop, versioned manifests with pinned tools, four autonomy patterns, shared contracts, and control points. Model selection lives in the LLM Blueprint. MCP is transport only.
How to build a robust evaluation framework across every AI system plane: data sources, offline and online modes, three scorers, specialized automation, and per-plane playbooks.
How to implement a Policy-Governed Agent Runtime. PEP/PDP enforcement, subject-action-resource-context contracts, audit replay, step-up, and boundary-specific playbooks.
Reference design for G.A.I.N MCP as transport: client, gateway, server registry. Tool contracts (manifests) live in the Agents Blueprint.