Most Multi-Agent AI Failures Stem From Poor Design, Not Model Limitations

A 2025 study by UC Berkeley researchers analyzed over 1,600 annotated traces across seven multi-agent AI frameworks and identified 14 distinct failure modes. The research found that performance gains from multi-agent systems on popular benchmarks are often minimal, with failures rooted in system design flaws, inter-agent misalignment, and inadequate task verification. Separately, Anthropic's internal research showed a multi-agent system outperformed a single agent by 90.2%, but the gains were largely attributed to parallel compute spending rather than collective reasoning. The study authors noted that targeted fixes like clearer role definitions and added verification steps improved success rates but were insufficient for reliable performance. Experts now suggest that multi-agent setups only justify their roughly 15-times higher token cost when tasks are genuinely parallelizable and the output value clearly outweighs the added expense.