Skip to main content
Back

Harbin Researchers Warn Mixed AI Agents Make Supply Chains Worse

Researchers pit DeepSeek and GPT agents against each other in a supply chain simulation to reveal how mixed AI reasoning levels amplify the bullwhip effect — and how information sharing fixes it.

Pro

Paper, Full Access
  • Core finding: Mixing LLMs of different reasoning sophistication across supply chain tiers worsens the bullwhip effect compared to homogeneous setups.
  • Setup: DeepSeek and GPT agents were assigned to retailer, wholesaler, distributor, and manufacturer roles in four cognitive heterogeneity scenarios.
  • Agent behavior: All LLM agents defaulted to myopic, self-interested ordering that amplified order variability upstream.
  • Key fix: Information sharing across tiers dramatically reduced order variance in all scenarios, even when agent reasoning capabilities were mismatched.
  • Practical implication: In real deployments with mixed-capability AI systems, structural transparency (shared demand/inventory data) compensates for reasoning gaps between agents.
  • Limitation: No code released; LLM response distributions diverge from human behavior, limiting direct generalization to real supply chains. Read the paper.

What happens when you staff a supply chain with AI agents of wildly different reasoning abilities? A new paper from researchers at Harbin Institute of Technology set out to answer exactly that, using LLMs as stand-ins for human decision-makers in a classic supply chain simulation. The results are a cautionary tale: cognitive diversity among AI agents does not naturally lead to better coordination. Instead, it tends to make things worse , unless you force the agents to share information.

The Bullwhip Problem, Briefly

The bullwhip effect is one of the most well-documented inefficiencies in supply chain management. The classic setup involves four agents , a retailer, a warehouse, a distributor, and a manufacturer , who must make independent replenishment decisions with limited information. The game is widely used to demonstrate the bullwhip effect, a phenomenon in which order variability increases as one moves upstream in the supply chain. A small uptick in consumer demand at the retail end can cause the manufacturer to wildly overproduce, because each intermediary over-orders to protect themselves.

The bullwhip effect arises from behavioral factors that lead to decision-making biases. Supply chain managers' decisions can be influenced by misinterpreting demand data, resulting in overly optimistic forecasts and confusion between forecasts and targets. The issue is not just technical , it is behavioral. The bullwhip effect is not just a supply chain problem; it is a people problem.

Supply chain process flow diagram showing product and order flows across manufacturer, distributor, wholesaler, and retailer

Traditional research on this problem relies on human lab experiments or stylized mathematical models. Both have hard limits: you can only run so many human trials, and mathematical models strip out the messy cognitive realism of actual decision-making. LLM-based simulation offers a third path , scalable, repeatable, and capable of exhibiting something resembling genuine behavioral biases.

The Setup: Mixing Brains Across the Chain

The paper introduces a Hierarchical Reasoning Framework to structure how agents at different supply chain tiers reason about their decisions. The key twist is