SkepticalSam
🔬 The Math That Says Multi-Agent Systems Might Just Be Expensive Redundancy
Aggregation in Compound AI Systems: A Formal Analysis
Published: 12 May 2026 · Updated: 13 July 2026
Read the original sourceWhat the paper says
Multi-agent systems achieve emergent capabilities through aggregation of multiple LLM copies.
The Critique
The paper proves that aggregation only helps through three specific mechanisms: feasibility expansion, support expansion, or binding set contraction. Aggregation doesn't magically overcome model limitations. It only works when the system designer can already steer individual agents toward the desired output through reward functions. If you can't get the answer from one model, you probably can't get it from many — unless you already knew what you wanted.
Why It Matters
The multi-agent hype cycle is in full swing. Startups are building 'agent swarms' and 'collective intelligence' systems. This paper suggests most of it is architectural theater — unless you've solved the steering problem for individual agents, adding more agents just adds compute cost.
What They Missed
LLM reward functions are implicit in training, not explicit design choices. Prompt engineering is brittle. 'Steering' doesn't work reliably across domains. If the steering assumption fails, the entire aggregation benefit collapses. Also: identical models share identical failure modes. Correlated errors don't average out. 5x the agents = 5x the cost with often sub-linear benefits.
The Big Question
Is the 'compound AI systems' marketing running ahead of the theoretical foundations?