Search papers, labs, and topics across Lattice.
This paper investigates the design of strategic buying agents that autonomously decide when to purchase products within a limited timeframe, leveraging price observations and beliefs about future price changes. The authors formulate the purchasing problem across three regimes鈥攕tationary, Bayesian, and robust鈥攄eveloping optimal purchase policies that are implemented as a policy menu. Evaluating these policies against real Amazon price histories reveals that while stationary and Bayesian policies yield competitive consumer surplus, the robust policy excels in worst-case scenarios, indicating that language models can effectively select among different purchasing strategies rather than making direct buy-or-wait decisions.
Language models can outperform traditional methods by effectively selecting optimal purchasing strategies rather than making direct buy-or-wait decisions.
Agentic AI is shifting online shopping from search toward delegated purchasing, where autonomous buying agents monitor markets and decide when to buy on a consumer's behalf. We study the design of such strategic buying agents, which must decide when to purchase within a finite shopping window, translating price observations, the remaining time horizon, and beliefs about future price changes into a purchase policy. We formulate this problem across three information regimes: stationary, Bayesian, and robust, and treat the resulting optimal policies as a policy menu for implementation. In the stationary regime, price adjustments follow a Poisson arrival process with a known post-adjustment price distribution; the optimal policy is a dynamic purchase-threshold rule, with the threshold governed by an ordinary differential equation. In the Bayesian regime, the adjustment intensity is known, but the price-adjustment distribution is uncertain; the optimal rule remains threshold-based, now depending on posterior beliefs, and we bound the value of knowing the true distribution. In the robust regime, the agent has only price bounds and seeks worst-case protection; randomized threshold policies achieve optimal competitive-ratio and minimax-regret guarantees. We evaluate the proposed policies on Amazon price histories from Keepa (367 items, 48,933 timestamped observations) and examine their integration into language-model buying agents. The stationary and Bayesian policies perform competitively on mean normalized consumer surplus despite their stylized assumptions, while the robust policy performs best at the distribution's 10th percentile. Results suggest language models are better suited to selecting among regimes and calibration samples than to making buy-or-wait decisions directly.
