Can multi-agent teams swap roles, coordination, and algorithms independently?

IMACS (Intelligent Multi-Agent Collaboration System) argues that multi-agent LLM frameworks "routinely entangle three logically distinct concerns: who is on the team (organization), how members align (coordination), and which algorithm fuses their work (collaboration protocol)." Its thesis is that these are orthogonal and should be implemented as independently swappable layers rather than baked together. Organizations are declared in YAML and grounded in Belbin's nine team roles (Belbin 2010); coordination is chosen from Mintzberg's five coordination mechanisms (Mintzberg 1979); RACI (Responsible/Accountable/Consulted/Informed) is a cross-cutting tag with a validated constraint that at most one role can be Accountable per deliverable; and the collaboration protocol — debate, voting, Mixture-of-Agents, blending, reflexion, plan-and-execute — is "a pluggable object behind a single interface" that six published algorithms are placed behind.

Formally, an organization is a tuple O = (R, μ, κ, ρ): a set of Belbin-typed roles R, a model binding μ from roles to a model registry, a single Mintzberg coordination mechanism κ, and a RACI assignment ρ subject to the single-Accountable constraint. A collaboration protocol is a function π that takes a task, an instantiated team, and a shared blackboard, and returns a result while posting every intermediate message to the blackboard. The paper states "a protocol is deliberately agnostic to which roles populate the team and to κ; it sees a pool of role-typed agents, never a fixed cast," and "an organization carries no commitment to any π" — the decoupling is architectural, not just conceptual, and is stated as a proposition (any layer can be swapped without touching the others).

This sits as a direct counterpoint to Does structured artifact sharing outperform conversational coordination?: MetaGPT and ChatDev, as IMACS's own related-work section notes, "bind roles to a fixed software-company pipeline" — the organizational theory (roles, handoffs) and the collaboration algorithm are fused into one implementation. IMACS's design-space argument is that this fusion is the problem: when roles and protocol are entangled, "swapping the aggregation algorithm means rewriting the team," organizational theory "cannot be applied as a first-class design surface," and algorithms "cannot be compared on an equal organizational footing, since each ships its own implicit team." It also parallels Can we automatically optimize both prompts and agent coordination? and Do single agents always hit organizational limits? in treating multi-agent structure as an explicit, manipulable layer rather than an emergent byproduct of prompting — though IMACS's layer is organizational theory (Belbin/Mintzberg/RACI) rather than a graph of nodes and edges.

The excerpt is a method and framework paper from its own authors, not an independent evaluation of IMACS against competing frameworks, and it does not establish that Belbin roles or Mintzberg coordination mechanisms are the "correct" or complete taxonomy for agent teams — only that treating some such taxonomy as a swappable, validated configuration layer enables controlled ablation that fused frameworks cannot do. The paper's own limitations section flags that its quantitative comparisons rest on a synthetic oracle calibrated to literature trends and small-N real-model runs (N=10 per cell), so the architectural decoupling claim is better supported than any specific performance ranking among the six protocols it hosts.