Reliability at the core: five ground rules for building an agent loop