Studies Find Multi-Agent Systems Amplify Errors by Up to 17.2x
Two independent 2026 studies — Stanford on thinking-budget controls and Google/MIT on coordination overhead — find multi-agent architectures systematically underperform. Tool-heavy workloads (16 tools) show single-agent coordination efficiency of 0.466 vs 0.074–0.234 for multi-agent: a 2x–6x penalty. Multi-agent appeared superior only when thinking budget was not held constant — a confound, not an architectural advantage.
Why It Matters
The empirical case against default multi-agent architecture is now published and peer-reviewed. Teams building production agentic systems should treat strong single-agent as the baseline, adding multi-agent only for genuinely decomposable parallel workloads.