Existing multi-agent software development frameworks such as ChatDev and MetaGPT organize agents around roles (product manager, developer, tester) within a single simulated organization. While effective for monolithic tasks, this role-playing paradigm breaks down in real-world polyglot systems where multiple independently-deployed services—each maintained by its own LLM agent—must coordinate across service boundaries. We present AgentNexus, a document exchange center that coordinates heterogeneous code agents at the service granularity rather than the role granularity. AgentNexus introduces three key ideas: (1) a versioned Markdown document store with publish-subscribe notification, enabling agents to detect and respond to cross-service changes; (2) an explicit lifecycle stage model that tracks each service’s development phase as a first-class entity, replacing ad-hoc role-playing with structured state transitions; and (3) a diff-aware update protocol that delivers structured change summaries alongside full document context, allowing downstream agents to perform targeted code modifications. We describe the architecture, implementation, and an initial deployment coordinating a backend search service and its frontend management console. Our results suggest that grounding multi-agent coordination in service-level document exchange—rather than simulated organizational roles—better reflects the structure of real software systems and reduces coordination overhead.
Longhui Tian (Tue,) studied this question.