MAR is reusable, project-agnostic execution infrastructure for AI applications. It owns the mechanics of running agents while leaving domain behavior to the applications that consume it.

The runtime provides typed contracts and extension points for orchestration, model routing, tools, policy, context assembly, memory, state, lifecycle hooks, delegation, diagnostics, budgets, retries, cancellation, and optional HTTP integration.

Applications own domain behavior; the runtime owns execution mechanics.

The kernel depends on narrow protocols instead of concrete providers.

Agent runs are the fundamental execution primitive, including delegated child runs.

Policy decisions, tool calls, memory writes, and lifecycle changes remain explicit and observable.

Runtime events are diagnostics, not automatically persisted memories.

Optional adapters isolate dependencies on external frameworks and services.