Do multi-agent LLM systems scale better than centralized hierarchies?
This reads the question as a head-to-head — do agents that coordinate as peers handle growth better than a top-down command structure — and the corpus suggests the real answer is that neither pure form scales, while a hybrid that fixes structure but frees roles wins.
This explores whether decentralized agent teams outscale top-down hierarchies — and the corpus reframes the contest as a false binary, because the most direct evidence says neither extreme wins. A 25,000-task experiment across eight models found that the best architecture was a hybrid: external ordering imposed from outside (the 'hierarchy' part) but with agents choosing their own roles internally. That setup beat centralized systems by 14% and fully autonomous, self-organizing systems by 44% Do self-organizing agent teams outperform rigid hierarchies?. So the lesson isn't 'distributed beats centralized' — it's that you want a skeleton of fixed structure with autonomy living inside it.
The deeper finding is that scale itself is the enemy for any multi-agent setup, regardless of how you organize it. Coordination degrades predictably as the network grows, because agents agree too late or adopt strategies without telling their neighbors, and they accept each other's information without verification — so errors propagate Why do multi-agent systems fail to coordinate at scale?. The same size-penalty shows up in consensus: groups fail not through subtle corruption but through liveness loss — timeouts and stalled convergence — and agreement gets worse as the group grows even with no bad actors present Can LLM agent groups reliably reach consensus together?. And the failures are characteristically LLM-shaped: role flipping, flake replies, infinite loops, and conversation drift, all because models lack a persistent goal and a stable sense of which role they're playing Why do autonomous LLM agents fail in predictable ways?.
More unsettling for the 'distributed scales better' thesis: agents don't seem to develop the social structure that would let a large system self-organize. On a platform with millions of interacting agents, they ignored feedback, showed no co-evolution, and never formed stable influence structures or shared memory — even though the memory and communication infrastructure existed Why don't AI agents develop social structure at scale?. Scale didn't produce emergent order; it produced noise.
There's also a sobering question of whether multi-agent gains are even real. One analysis found 80% of multi-agent performance variance comes simply from total token budget, not from coordination intelligence — meaning a lot of what looks like 'the team is smarter' is just 'the team spent more compute' How does test-time scaling work at the agent level?. Relatedly, the whole multi-agent advantage shrinks as single agents get stronger, with single-agent systems winning outright in many cases once node-level bottlenecks, edge-level overload, and error propagation are accounted for When do multi-agent systems actually outperform single agents?. And a single LLM using structured branching prompts can replicate multi-agent debate dynamics without spinning up multiple instances at all Can branching prompts replicate what multi-agent systems do?.
The thread tying these together is that scaling reliability comes from structure outside the model, not from the topology of the agents. Reliable systems externalize memory, skills, and protocols into a harness layer rather than hoping coordination emerges from more agents Where does agent reliability actually come from?. So the honest answer to the question is: multi-agent systems don't inherently scale better than hierarchies — both degrade with size — and the thing that actually scales is a designed harness with fixed structure and bounded autonomy, often built from cheaper specialized models rather than ever-larger swarms Can small language models handle most agent tasks?.
Sources 10 notes
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.
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
Across hundreds of simulations, LLM-agent groups frequently fail to reach valid agreement due to timeouts and stalled convergence rather than subtle value corruption. Agreement degrades with group size even without Byzantine agents present.
Research identifies role flipping, flake replies, infinite loops, and conversation deviation as LLM-specific failures in multi-agent cooperation. These occur because LLMs lack persistent goal representation and stable role identity.
A study of Moltbook, a platform with millions of interacting agents, found that agents ignore feedback, show no adaptive co-evolution, and never develop stable influence structures or shared social memory—despite having memory infrastructure and communication channels.
Research shows 80% of multi-agent performance variance comes from token budget, not coordination intelligence. LatentMAS and shared-KV-cache approaches offer ways to decouple performance gains from token costs.
Empirical analysis shows MAS performance gaps narrow with stronger models, with SAS outperforming in many cases. Three formal defect types—node-level bottlenecks, edge-level overwhelm, and path-level error propagation—explain when single agents win.
Research shows single LLMs using dynamic persona simulation achieve multi-agent cognitive synergy without multiple model instances. Solo Performance Prompting validates that structured prompting techniques map directly to multi-agent debate architectures, enabling equivalent outcomes through structural equivalence.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
SLMs handle the repetitive, well-defined language tasks that constitute most agent work at 10–30× lower cost than LLMs, making heterogeneous architectures (SLMs by default, LLMs selective) the economically rational design pattern.