What structural constraints does topology impose on role and LLM assignment?
This explores how the wiring of a multi-agent system — who talks to whom, in what order — shapes the downstream choices of which job each agent does and which model powers it.
This explores how the *shape* of a multi-agent system (its communication topology) constrains two other decisions: what role each agent plays and which LLM gets assigned to it. The corpus suggests these aren't independent knobs you can tune one at a time — topology, roles, and model choice are entangled, and the best systems decide them jointly rather than in sequence.
The clearest answer comes from MasRouter What decisions must multi-agent routing systems optimize simultaneously?, which treats routing as four decisions that must be optimized *at once*: the collaboration topology, how many agents there are, what role each takes, and which model powers each one. The order matters — once you fix how agents are connected, you've already narrowed which roles make sense and how expensive the model assignment can be. A cascaded controller that respects this dependency beat single-model routing by 3.5% accuracy while cutting costs nearly in half. The lesson is that topology is the upstream constraint: it sets the stage on which roles and models can be cast.
But a striking counter-finding complicates the picture. A 25,000-task experiment Do self-organizing agent teams outperform rigid hierarchies? found that the winning recipe was *fixed external structure with autonomous internal roles* — agents handed a rigid ordering but left free to invent their own specializations (and to bow out when incompetent) outperformed both fully centralized and fully autonomous designs. So topology should be imposed, but role assignment should be released. The structure constrains the workflow; the roles emerge inside it. This is a different kind of constraint than MasRouter's joint optimization — it says the *right* constraint is to fix only the scaffolding and let roles self-organize.
There's a deeper, almost cautionary thread here about what 'topology' even buys you. Work on whether LLMs can use graph structure Can language models actually use graph structure information? found that models recognize graphs as a *category* but largely ignore the actual connections — shuffling the edges barely changed performance. If models are weak at genuinely reasoning over relational structure, then topology constraints in agent systems may be doing their work externally (in the orchestration harness) rather than being internalized by the agents themselves. That reframes the whole question: topology constrains role/model assignment from the *outside*, as an engineered control flow, not because agents perceive the graph they live in.
That external-harness view connects to two adjacent ideas worth following. LLM Programs Can algorithms control LLM reasoning better than LLMs alone? show that explicit algorithmic control flow — deciding in advance which step sees which context — can decompose hard tasks better than letting the model freelance; that's topology-as-information-hiding, and it directly shapes what each call's 'role' is allowed to be. And the large-action-models work Can you turn an LLM into an agent by just fine-tuning? argues the surrounding system, not the model weights, determines whether behavior is grounded — another vote that structure is the binding constraint. If you want to go deeper on where the real leverage sits, What makes a research domain suitable for autonomous optimization? makes the parallel claim for whole domains: the bottleneck is environmental structure, not model power.
Sources 6 notes
MasRouter shows that routing in multi-agent systems must jointly optimize collaboration topology, agent count, role allocation, and per-agent LLM assignment through a cascaded controller. This unified approach surpasses single-model routing by 3.51% accuracy while cutting HumanEval costs by 49%.
A 25,000-task experiment across 8 models and multiple agent counts showed that sequential protocols with external ordering but internal role selection outperform centralized systems by 14% and fully autonomous systems by 44%. Agents spontaneously invented specialized roles and self-abstained when incompetent.
LLMs develop attention shifts toward node tokens after training, but randomly shuffled topology barely affects performance. Models treat graph data as a category to recognize rather than as structured relationships to use.
LLM Programs embed LLMs within explicit algorithms that manage control flow and state, presenting only step-specific context to each LLM call. This information hiding addresses capability and context window limits while treating complex reasoning as modular, debuggable sub-tasks.
Converting LLMs to action-capable systems requires four distinct stages: curating action-environment-user datasets, training for action grounding, integrating agent infrastructure with memory and tools, and rigorous safety evaluation. The surrounding system and harness determine whether actions are grounded or hallucinated.
Autonomous research pipelines require immediate scalar metrics, modular architecture, fast iteration cycles, and version control. Domains lacking any property resist autoresearch regardless of LLM capability, because the bottleneck is environmental structure, not model power.