Apple finds natural language bottlenecks multi-agent reasoning

Forcing AI agents to talk to one another in conversational human prose is an architectural anti-pattern that sabotages reasoning fidelity.

• –Generation is the primary failure mode: Over 73% of communication breakdowns occur during text generation by the sender, proving LLMs struggle to faithfully serialize complex internal logic into sequential words.
• –Structured protocols beat natural language: Multi-agent frameworks relying on typed graphs, ASTs, or explicit schemas prevent translation drift and eliminate serialization noise.
• –Swarm scaling limits: Adding more conversational agents to a problem exacerbates compounding errors rather than generating emergent collaborative intelligence.
• –Shift to formal agent IPC: Future multi-agent systems will inevitably move away from natural language prompts toward deterministic inter-process communication protocols.