Normalize the thinking-token budget and the multi-hop reasoning edge people credit to multi-agent orchestration stops reliably showing up 1. That one control changes what the swarm architecture is. The case studies that sell fan-out as a reasoning upgrade almost never report the token budget each side spent, and when a study does hold it constant, the reported multi-agent advantages are better explained by unaccounted computation and context effects than by anything the architecture does 1. So the honest reading is evidential, not causal: on current evidence the reasoning premium is confounded with compute you could have handed to one agent, and the coordination layer you bought instead is not free.
The comparison was never fair
The number that should make you suspicious is the one most write-ups omit: how many reasoning tokens each side actually spent. A five-agent debate that runs three rounds is not one model’s worth of thinking, it is closer to fifteen, plus the aggregate context each agent accumulates. Compare that against a single agent capped at one pass and the swarm wins, but you have not learned that orchestration reasons better. You have learned that more inference compute reasons better, which was never in dispute.
Hold the budget constant and the gap closes. In the equal-budget comparison, for multi-hop reasoning tasks the reported advantages of multi-agent systems are better explained by unaccounted computation and context effects than by inherent architectural benefits 1. Read that carefully, because the claim is narrower than the headline it refutes. It does not say multi-agent is worse. It says that once the confound is removed, the reasoning-quality win people attribute to the architecture does not reliably survive, and the wild gains reported in the wild are best read as an artifact of the extra compute fan-out silently spends.
There is a real steelman here, and I want to answer it head on. Token-normalization equalizes how much compute each side gets, but multi-agent changes more than one variable at a time: role-diverse prompting, independent sampling, debate and cross-checking dynamics. A single benchmark showing parity at equal tokens is consistent with “no architectural benefit” and also with “a real benefit this particular harness did not exercise.” I take that seriously enough to state the claim as falsifiable rather than proven: a budget-normalized, multi-hop benchmark where multi-agent reliably beats a single agent at equal thinking-token budget would refute it directly, and that experiment is buildable. Until it lands, the evidence points the other way, and the burden sits with the pitch that sells orchestration as a reasoning upgrade.
What the orchestration layer actually costs
The compute confound is only half the argument, and it is the fragile half, since it leans on one line of evidence. The other half is independently sourced and it is about what you pay to run the swarm at all. Coordination is not incidental overhead you tune away; its failure modes form a structured, recurring taxonomy. MAST catalogs 14 failure modes across three root categories, specification ambiguity, coordination breakdowns, and verification gaps, validated across more than 1,600 execution traces at NeurIPS 2025 2.
Make one of those concrete. A specification-ambiguity failure is the handoff where agent A finishes its slice against an interpretation of the task that agent B never shared, and neither notices, because no step in the pipeline is responsible for reconciling the two readings. Nothing crashes. The system produces a confident, internally coherent, wrong answer, and the coordination that was supposed to be the architecture’s strength is exactly where it breaks. That these span 14 recurring modes across more than 1,600 traces is the point: they are structural properties of putting a coordination layer between the model and the task, not bugs a better prompt retires.
When multi-agent still earns its keep
None of this makes multi-agent a mistake. It makes the reason you reach for it matter. There are two reasons that survive scrutiny and one that does not. Genuine wall-clock parallelism survives: if three sub-problems are independent, running them concurrently is faster in real time whatever the token accounting says. Context isolation survives: keeping a sub-problem’s messy intermediate state out of another agent’s context window is a real engineering benefit, especially as context windows fill and attention degrades. The reasoning-quality upgrade does not survive, at least not on current evidence.
That gives a clean decision rule. Reach for multi-agent when the value is structural, when you want parallel wall-clock speed or you want to quarantine context. Do not reach for it expecting the orchestration itself to reason better than one well-budgeted agent, and treat any vendor benchmark that claims otherwise as suspect until it reports the thinking-token budget on both sides.
One honest caveat applies to that carve-out. “Use it for context isolation” can become an escape hatch: reclassify any future multi-agent reasoning win as really being isolation and the reasoning claim becomes unfalsifiable. I do not want that move. The claim is specifically about reasoning-quality gains under an equal budget, so a win that survives budget-normalization refutes it and does not get relabeled. The coordination-tax leg 2 stands on its own source regardless, so even if the compute-confound finding is overturned tomorrow, the argument that the orchestration layer fails in structural, documented ways does not fall with it.