Why Single Agents Beat Multi-Agent Systems at Equal Token Budgets

Short summary

Stanford research (Tran & Kiela, arXiv 2604.02460) demonstrates single-agent LLMs outperform multi-agent systems when reasoning-token budgets are held equal, consistently across Qwen3, DeepSeek-R1, and Gemini 2.5. Prior benchmarks gave multi-agent systems 2–4x more tokens, biasing architecture decisions across the industry. Single agents with explicit step-by-step reasoning recover most multi-agent benefits without orchestration complexity.

• •Single agents beat multi-agent systems at equal token budgets across all three tested model families
• •Information loss at each agent handoff (Data Processing Inequality) explains the performance gap
• •Year of prior benchmarks unfairly gave multi-agent systems 2–4x more reasoning tokens, skewing industry decisions
• •Single agents with explicit pre-answer reasoning prompts recover multi-agent collaboration benefits
• •Decision framework: use single agent for reasoning tasks; multi-agent only for context fragmentation or genuinely higher budgets

Generated with AI, which can make mistakes.