Can separating structure and function improve multiagent organization design?

Jomi Fred Hübner, Jaime Simão Sichman, and Olivier Boissier, writing at SBIA 2002, argue that an organizational specification (OS) for a multiagent system should hold three parts rather than one: "An MOISE+ OS is formed by a Structural Specification (SS), a Functional Specification (FS), and a Deontic Specification (DS)." The structural specification is roles, the relations among them, and groups; the functional specification is the global plans (missions) that achieve the system's purpose; the deontic specification is what connects the two, stating which role is obliged or permitted to commit to which mission. Their stated objective is "an organization centered model where the first two dimensions can be specified almost independently of each other and after properly linked by the deontic dimension."

The paper's own reasoning for the split is a two-sided failure mode. "If only the functional dimension is specified, the organization has nothing to 'tell' the agents when no plan can be performed." "If only the organizational structure is specified, the agents have to reason for a global plan every time they want to play together" and "the plans developed for a problem are lost, since there is no organizational memory to store these plans." The authors illustrate with a soccer team: structure is the defense and attack groups and their roles; function is the rehearsed plays, "predefined plans that [were] already validated" and so do not need to be re-derived from scratch each match. Formally, structure narrows an agent's possible behaviors toward the set of behaviors matching the system's purpose, but getting the rest of the way there is left to validated plans supplied by the functional dimension — the deontic specification is what says which role owes which plan.

The deontic specification is a 2002, pre-LLM answer to a problem now being named again in agent-safety terms: What makes delegation work beyond just splitting tasks? treats authority transfer and accountability as properties a delegation contract must specify per task; MOISE+'s deontic relation is the organization-level slot that contract fills — declaring, once, which role is permitted or obliged to take on which mission, rather than negotiating it per delegation event. Likewise, MOISE+'s structural "authority link," which the paper defines as letting agents of a source role "control" agents of a destination role, states exactly the kind of authorization boundary that Why do agents fail at identity verification and authorization? finds missing from deployed LLM-agent systems — but MOISE+ assumes the boundary is declared by a designer in advance and followed, not negotiated, impersonated, or verified cryptographically at runtime, which is precisely the gap that NIST's initiative is chasing in today's systems.

The excerpt specifies a soccer-team example and gestures at an unpresented B2B case; it offers no evidence, from LLM agents or otherwise, that declaring structure, function, and deontic relations this way changes what agents actually do, only a formal account of what such a specification would permit. The authors also concede the model's own cost: an organization that constrains behavior toward its purpose can make "the MAS stiffer," so that "if the environment changes, the current set of allowed organizational behaviors may not fit the global purpose anymore," and they call a reorganization process "mandatory" to recover — governance by fixed structure buys predictability at the price of a flexibility problem the paper leaves unsolved.