Run workers and AgentRuntime.serve() processes as long-lived services. Scale by
adding worker instances and use task domains or queue limits to isolate workloads.
On shutdown, stop task handlers and runtimes so in-flight tasks can be redelivered
instead of being abandoned.
Do not construct a new runtime per web request. Reuse clients and runtimes for the application lifetime; use reliability for timeout and retry policy.