You don't need different AI models to get a real division of labor — identical agents will self-organize into specialists.

This reads the question as: does splitting a team of agents into roles require the models themselves to be differentiated, and the corpus gives a surprising answer — differentiation matters far less than *structure*. The strongest result is that the best-performing teams don't hand-design who does what; they impose a fixed sequence but let agents pick their own roles, and agents will spontaneously invent specializations and even abstain when they judge themselves incompetent for a subtask Do self-organizing agent teams outperform rigid hierarchies?. So role allocation can emerge from position and self-assessment rather than from the models being intrinsically distinct.

Where model differentiation *does* pay off is economic, not cognitive. Most agent work is repetitive, well-defined language processing that small models handle at a fraction of the cost, which makes the rational design a heterogeneous team — cheap small models by default, expensive large ones called in selectively Can small language models handle most agent tasks?. Here role and model tier are deliberately matched: you differentiate the models precisely so you can allocate the boring bulk to the cheap ones. And capability tier maps onto roles non-monotonically — the ability to improve a shared harness is flat across model sizes, but the ability to *benefit* peaks in the middle, so your strongest model isn't automatically your best role-filler Do stronger models always evolve harnesses better?.

The more interesting twist is that roles help even when the models are identical. A leader-follower debate — one agent proposes interpretations, two others challenge with rotating turns — lifts a small 7B model's ambiguity detection dramatically, and the gain comes from the role *structure* and forced rotation, not from any model being smarter than another Can structured debate roles help small models detect ambiguity?. Differentiation is manufactured by the seating chart, not inherited from the weights. The flip side of position-as-role is danger: in a workflow, whoever occupies a high-influence, dependency-converging slot can amplify signals — including malicious ones — far downstream, so the same positional power that makes role allocation work also concentrates risk How does workflow position shape attack propagation in multi-agent systems?.

A cleaner way to think about allocation is to stop treating the model as the unit at all. Reliable agents externalize memory, skills, and protocols into a shared harness layer rather than leaning on any single model's raw scale agent-reliability-comes-from-externalizing-cognitive-burdens-into-system-structures, and teams can measure each member's actual contribution and deactivate the freeloaders at inference time Can multi-agent teams automatically remove their weakest members?. Roles get allocated by measured usefulness, and low-contributors get pruned regardless of what model backs them.

So the reader's hidden lesson: role allocation depends surprisingly little on models being different, and surprisingly much on structure, position, and self-abstention. The one place differentiation genuinely earns its keep is cost — matching cheap models to routine roles — not competence. Worth noting the corpus is thin on the specific case of *asymmetric-information* role assignment: models look socially competent only when one model secretly controls everyone, and fail once agents hold genuinely private information Why do LLMs fail when simulating agents with private information? — which suggests differentiating agents by *what they know* is a harder and less-solved axis than differentiating them by how big they are.