Single-Agent vs Multi-Agent Systems

Equal budgets change the result

Tran and Kiela (arXiv, April 2026) compared a single agent against five multi-agent designs (sequential, subtask-parallel, parallel-roles, debate and ensemble) on the FRAMES and MuSiQue multi-hop benchmarks, across Qwen3, DeepSeek-R1-Distill-Llama and Gemini 2.5 models, at thinking budgets from 100 to 10,000 tokens. With reasoning tokens held constant, the single agent "consistently match[ed] or outperform[ed]" the multi-agent systems. Their explanation uses the Data Processing Inequality: passing information between agents cannot add information and can lose it, so a single reasoner with full context is the more information-efficient use of a fixed budget.

The paper also explains some earlier contrary results. Multi-agent pipelines usually consume more total compute, and the authors found that API-level budget controls, particularly on Gemini 2.5, let multi-agent runs surface more reasoning than the nominal budget implied. The scope is limited: text-only multi-hop reasoning, no tools, and the authors place tool use and safety constraints outside the study.

Generated architectures versus designed ones

Jwalapuram et al. ("The Illusion of Multi-Agent Advantage", June 2026) tested six frameworks that generate multi-agent architectures automatically, including ADAS, AFlow, MaAS and DyLAN, against a single-agent baseline of chain-of-thought with self-consistency. Across GPQA-Diamond, HLE-Maths, SWE-Bench Lite and BrowseComp-Plus, the generated systems underperformed the baseline while costing up to ten times as much. Inspection showed what the authors call architectural bloat: agents that reached unanimous agreement in roughly 70% of GPT-4o cases, and, for one framework, half of the discovered workflows amounting to one prompt repeated three times.

The same paper contains the strongest evidence for multi-agent design. On a synthetic benchmark built with explicit sub-tasks, separated context and room for parallel work, an expert-designed system reached 96.51% against 56.97% for the single-agent baseline, at comparable cost. The lesson is specific: structure helps when the task has that structure and a person has matched the architecture to it. It is one synthetic benchmark from one paper.

What practitioners report

Vendor accounts line up with this split. Anthropic's June 2025 description of its research system reports a 90.2% improvement for a lead agent with subagents over a single agent on an internal breadth-first research evaluation, and also that token usage alone explained 80% of performance variance on BrowseComp. That is the compute confound the academic studies control for. Anthropic states that multi-agent systems suit tasks with heavy parallelization and information exceeding one context window, and are a poor fit when agents need shared context. Cognition's Walden Yan argued the same month that parallel subagents make conflicting implicit decisions, and recommended a single-threaded agent with context compression for long tasks.

Context is the real variable

Tran and Kiela's degradation experiments point to the deciding factor. Multi-agent designs became competitive when the single agent's context was corrupted enough that one reasoning trajectory could no longer separate relevant from misleading material. Subagents, in other words, are a context-management device. They pay off when isolating work in fresh contexts protects the lead agent more than the handoff costs it. Alternatives such as context compaction address the same pressure inside one agent.