Count the tokens first: single agents hold up against swarms on multi-hop reasoning

Dat Tran, Douwe Kiela · 2 authors · Stanford University · 2026 · arXiv:2604.02460

Does the swarm still win once you count every thinking token?

Opening

Splitting a hard question across several agents is the reflex pitch in agent engineering: more agents, more capability. What the pitch rarely counts is the thinking those agents burn. Test-time computation confounds almost every published comparison, because a swarm that wins may simply have spent more tokens than the single model it beat. For anyone sizing a Kanban pipeline or quoting a client a monthly token budget, that distinction is the whole business. This paper normalizes the bill and asks the question plainly: when both sides get the same reasoning-token budget, does the multi-agent architecture still pull ahead?

What the paper does

The authors open with an information-theoretic argument grounded in the Data Processing Inequality: under a fixed reasoning-token budget and with perfect context utilization, a single-agent system (SAS) is more information-efficient than a multi-agent system (MAS). That argument makes two predictions, namely that MAS becomes competitive when one agent’s context utilization degrades, or when more compute is spent. They then run a controlled empirical study across three model families, Qwen3, DeepSeek-R1-Distill-Llama, and Gemini 2.5, comparing SAS against multiple MAS architectures under matched budgets on multi-hop reasoning tasks. Beyond headline scores, the study audits the measurement itself, identifying artifacts in API-based budget control and in standard benchmarks that can inflate apparent MAS gains.

Findings

Normalized computation — Once computation is normalized, the extra heads are not buying capability, they are buying spend: the single agent matches or beats the multi-agent setup.

When computation is normalized, single-agent systems (SAS) can match or outperform MAS

Held-constant tokens — The result is not a one-off across a single model: with reasoning tokens pinned, SAS consistently matches or outperforms MAS on multi-hop reasoning across the three model families.

We find that SAS consistently match or outperform MAS on multi-hop reasoning tasks when reasoning tokens are held constant.

The real explanation — Most headline multi-agent wins trace back to unaccounted computation and context effects rather than to architecture, which also means a benchmark can quietly flatter a swarm.

many reported advantages of multi-agent systems are better explained by unaccounted computation and context effects rather than inherent architectural benefits

Caveat

The limitation the authors state is a boundary on their own conclusion: multi-agent systems become competitive when a single agent’s effective context utilization is degraded, or when more compute is expended. The single-agent advantage is therefore conditional, not universal. It holds where one agent’s context stays intact and budgets are genuinely matched, and weakens exactly where it stops holding.

What this means for 5ac

The default 5ac delegation shape should be one capable agent with a clear task, not a fan-out, until there is evidence the fan-out earns its tokens. When a Kanban card does need children, the card body should carry a budget: how far the parent may push delegation before cost per task stops beating the single-model route. Route ordinary drafting cards to the cheap model and reserve multi-agent work for cards whose context is genuinely fragmented, the condition under which this paper says MAS becomes competitive. The sales line changes with it: 5ac sells test-time cost control, not agent headcount.

Source