ACM paper explains why multi-agent AI can’t truly talk

The author argues that multi-agent AI systems do not actually converse, but instead execute brittle state-machine handoffs across remote procedure calls. Treating these exchanges as human-like dialogue expands token overhead exponentially, multiplies inference costs, and creates compounding hallucinations and security blind spots. Enterprise architectures should replace conversational abstractions with explicit state management, typed schemas, and deterministic guardrails. Open article →