A new paper from Stanford (Tran & Kiela, arXiv 2604.02460) tested single-agent vs multi-agent systems with identical thinking-token budgets — and the multi-agent advantage disappears.
The hidden variable
Every "multi-agent wins" benchmark you've read about let the multi-agent system spend 2–4x more reasoning tokens than the single agent. Pin the budget — the advantage disappears.
Methodology
- 3 model families: Qwen3, DeepSeek-R1-Distill-Llama, Gemini 2.5
- 4 token budgets from 100 to 10,000
- 2 multi-hop reasoning datasets (FRAMES, MuSiQue 4-hop)
- 5 MAS architectures vs SAS
Result: Single agent produced higher accuracy AND consumed less compute, across all three families.
The Gemini 2.5 API artifact
Paper exposes that Gemini 2.5 API doesn't actually enforce thinking-token caps reliably. MAS surfaces more visible thinking under the same requested budget because of multi-call structure. A year of benchmarks ran on a broken ruler.
Why it works this way
Data Processing Inequality (Shannon 1948, Fano 1952): every handoff between agents is a processing step that can only preserve or lose information — never add. Each agent receives a summary of what the previous did. Summary is lossy by definition.
Decision boundary
- Reasoning depth (multi-hop logic, chained inference) → single agent
- Context fragmentation (long heterogeneous docs, parallel sub-tasks) → multi-agent
Most multi-agent deployments are reasoning-depth problems mislabeled as fragmentation problems.
Practical rule
Before building the next multi-agent system, run the cheap experiment: take the task you'd give to 4 agents, give it to 1 with equal total token budget and explicit pre-answer analysis. If it matches — you don't need multi-agent.
Paper: arXiv 2604.02460
