SwarmSpec: A Declarative Architectural Contract for Governable LLM Multi-Agent Systems

LLM multi-agent systems require descriptions of coordination, capabilities, external interactions, and constraints on actions. Existing declarative representations already address important parts of this problem; I examine the relationships that should be made explicit across these contracts. This design paper presents SwarmSpec 1.0.0 as an architectural intermediate representation. Its model connects reusable agents, graph activations, state, models, tools and skills, interaction channels, policies, provenance, evaluation criteria, and artifacts. A catalog of 25 patterns maps coordination structures to required primitives and roles. I propose three contributions: this organization of the architectural contract; a conformance framework separating pre-execution checks, runtime obligations, and preservation requirements for execution targets; and a completion reserve assigned to identified nodes, paths, and artifacts, with explicit permitted uses and exclusions. Related work distinguishes these proposals from established approaches to portability, skills, and finalization reserves. A technical memorandum case illustrates their connections, while a conditional invariant clarifies the scope of budget protection. The paper contributes a model and an evaluation protocol, rather than experimental results on cross-runtime portability or task quality.