Do language-model agents amplify the bullwhip effect?
Agent-based modelling is rising in the operations research journals, and firms are starting to let AI agents place orders. The bullwhip effect, where small swings in demand grow larger up the supply chain, appears on only 9 recent papers, and just one of them carries an agent or AI keyword. Do AI ordering agents dampen the bullwhip, or amplify it when they all react to the same signals?
What prompted this
In the operations research journals, "agent-based modeling" had 6 papers in July–September 2026 against about 3 expected (see the operations research dashboard). Since October 2025, "agent-based modeling" appears on 81 papers (40 in economics journals, 16 in operations research) and "LLM agents" on 44, while "bullwhip effect" appears on 9, only one of which carries a language-model, agent or AI keyword.
Questions
- In the classic beer distribution game, do language-model agents order more steadily than human players, or overreact?
- What happens when every tier of the supply chain uses the same model?
- Does sharing point-of-sale data with the agents remove the bullwhip, as it does for people?
Possible approach
Run the beer game with a language-model agent at each tier, varying the information the agents share and whether all tiers use the same model, and measure how much the variance of orders grows from tier to tier.
Next step
Implement a four-tier beer game with language-model agents and compare the order variance at each tier with published results for human players.