Multi-Agent AI Handoffs Can Inflate Token Costs Up to 15x, Anthropic Finds

Anthropic reported in June 2025 that multi-agent systems consume roughly 15 times more tokens than single-agent chat interactions, though they also scored 90.2% higher on an internal research evaluation. The cost spike occurs because each agent in a handoff workflow typically receives duplicated context — including instructions, prior messages, retrieved content, and tool outputs — rather than a compact task packet. In LangGraph-based systems, this token amplification is compounded when loops or repeated handoffs resend the same payload across multiple graph nodes. Developers are advised to use tracing tools like LangSmith to measure token usage at each node before optimizing, rather than assuming multi-agent architectures are inherently wasteful. Experts recommend first evaluating whether a task genuinely requires multi-agent coordination or whether a single, well-equipped agent would be sufficient and more cost-effective.