DAIR.AI warns multi-subagent architectures collapse

The multi-agent narrative where dozens of autonomous agents collaborate seamlessly remains largely marketing fiction, as frontier models lack the coordination stability and mutual trust needed for complex multi-agent delegation.

• –Single orchestrator-to-executor delegation is currently the practical ceiling; adding secondary subagents tends to introduce exponential coordination failure and token waste.
• –Subagents excel primarily at read-only tasks like parallel search, literature review, and code analysis, where isolated execution cannot corrupt shared state.
• –Harness engineering triumphs over agent sprawl: durable shared state, structured scratchpads, and explicit message boards offer far better reliability than dynamic multi-agent graphs.
• –Frontier models struggle with diverse distributed contexts; until models are explicitly trained for inter-agent coordination protocols, deep multi-agent trees remain cost-prohibitive.