MIT: AI Agents in Supply Chains Create Bullwhip Effect Despite Outperforming Humans
MIT researchers using the Beer Game supply-chain simulation find that reasoning-model AI agents reduce costs by up to 67% compared to human teams — but introduce a new failure mode: decision-variance amplification across echelons, dubbed the agent bullwhip effect. Repeated sampling cannot resolve it because the amplification is inherent to multi-agent coordination with information delays. GRPO post-training with system-level (not agent-level) rewards successfully curtails the effect.
Why It Matters
The agent bullwhip effect generalizes beyond supply chains to any multi-agent pipeline where agents at different layers coordinate under information delay — including orchestrated agentic workflows. This is an empirical warning that optimizing each agent locally does not guarantee system-level stability.