This paper introduces Agent Rooms as a conceptual framework for persistent, human-governed multi-agent collaboration. As LLM-based agents become increasingly capable of planning, using tools, coordinating with other agents, and producing complex artifacts, existing abstractions such as chats, crews, workflows, and simulations often remain insufficient to describe the durable organizational spaces where long-running agentic work takes place. An Agent Room is defined as a persistent collaboration container that brings together human and AI participants, roles, goals, artifacts, memory, policies, communication protocols, audit events, and governance boundaries. The paper argues that multi-agent collaboration should not be treated merely as message exchange or task execution, but as a governed socio-technical process in which artifacts are created, revised, reviewed, approved, and preserved over time. The contribution includes a formal definition of Agent Rooms, a lifecycle model, a communication protocol architecture, a context-engineering and memory-governance model, a failure-mode analysis, and an evaluation protocol. The paper also presents Campaign Builder as a reference prototype that instantiates several core mechanisms of the framework, including planner-mediated collaboration, structured handoffs, persistent transcripts, artifact blackboards, audit events, bounded agent mentions, recursive sub-rooms, governed context selection, quality gates, and observable evaluation hooks. The prototype is presented as an implemented design reference rather than definitive empirical validation. The paper’s main goal is to define a clearer unit of analysis and design for auditable, artifact-producing, human-governed multi-agent systems.