Automated multi‑agent organization design is shifting from handcrafted prompts and workflows toward automated search over agent programs and interaction structures. Multi‑agent systems are gaining attention, yet a core question remains: when does an automatically generated multi‑agent structure exhibit a functionally meaningful division of responsibility? We argue that design should move beyond finding effective architectures to identifying which responsibilities are functionally necessary, how they should be allocated, and when they need to change. To this end we introduce Automated Multi‑Agent Organization Design (AMOD), treating the organization itself as a primary design object and explicitly modeling team composition, functional responsibilities, coordination rules, and collective decision mechanisms. AMOD follows an execution‑driven design loop: for each task an automated designer instantiates an organization by assigning responsibilities and coordination rules; after execution the task outcomes, intermediate artifacts, and interaction traces are examined to detect missing, duplicated, or ineffective responsibilities, and the organization is revised accordingly. We also propose testing discovered responsibilities by removing them, replacing them with generic roles, or altering their assignment under matched resource budgets. This framework provides a principled foundation for evidence‑based, adaptive multi‑agent organization design.
Review