Declarative Dialog Flows for Multi-Tenant Conversational Agents: Design Lessons from ARC

Abstract

ARC is a multi-tenant platform on which small businesses build conversational agents for chat, messaging and telephone. Authors draw an agent as a typed node graph covering conversation, logic, data handling and external capabilities, and a server-side runtime executes it together with intent classification, document retrieval and a catalogue of more than 200 integration connectors. This report describes three design choices and the problems that motivated them. First, the flow runtime is a pure step function that advances one user turn and expresses every side effect as an explicit request, so the same graph runs identically in an authoring simulator and in production. Second, agents installed from a shared template now follow that template automatically, using a content-hash version and a conditional write so an upgrade happens exactly once and preserves each business's own integration choices. Third, the variables that automations pass into conversations are governed by one declaration registry checked by a parity test, after the same declaration drift recurred three times and once caused an agent to speak template syntax to a caller. We report the operational measurements we have and state what remains unvalidated.

Keywords: conversational agents, dialog management, low-code platforms, multi-tenancy, software architecture