Nebula: Decoupling Agent Identity from Runtime in Multi-Agent LLM Systems

As LLM-based agents move from single-session tools to persistent, collaborative systems, a fundamental architectural question arises: who owns the agent's state? In most existing frameworks, the execution runtime owns the conversation history, and the orchestration layer has limited visibility into or control over individual messages. This paper describes Nebula, an open-source multi-agent platform that inverts this relationship by storing all agent state — messages, skills, knowledge, credentials, and memory — in a normalized relational database, separate from any execution runtime. We formalize this as a soul/body separation, where the soul (persistent identity) is composed into context-appropriate prompts and injected into disposable runtime processes (bodies) at each execution. We describe the architectural patterns this separation enables and discuss the trade-offs involved. Source code: https://github.com/reforia/Nebula