Strategic Buying Agents

Strategic Buying Agents: Optimizing Purchase Timing Under Price Uncertainty

When you ask an AI to buy something for you, it needs to decide exactly when to hit the "buy" button. This paper creates mathematical rules for these AI agents. They help agents decide whether to buy an item now or wait for a potential price drop, even when they aren't sure how often prices will change.

The rise of agentic AI is shifting online shopping from manual search toward delegated purchasing. In this new paradigm, autonomous agents monitor markets and execute transactions on behalf of consumers. However, a fundamental problem remains. How should an agent decide if a current price is "good enough"? Does the benefit of waiting outweigh the risk of missing the deal entirely?

The bottleneck of delegated commerce

Current e-commerce interactions are largely reactive. Platforms provide search tools and recommendations, but humans remain the primary decision-makers. Emerging web-agent systems can navigate interfaces and execute checkouts. However, they lack a formal logic for temporal decision-making (making decisions based on time). An agent might know how to click "add to cart." It does not inherently know how to weigh a \$379 price tag against a \$400 budget when a deadline is looming .

Existing approaches often fall into two traps. They are either too impulsive, buying immediately and missing discounts. Or they are too passive, waiting for a "perfect" price that never arrives before the deadline. Without a mathematical framework to handle price volatility, an autonomous buyer is essentially gambling. The authors argue the core problem is translating price observations and time horizons into a rigorous purchase policy.

Three regimes of market uncertainty

The authors propose a layered architecture to solve this. They separate the "formulation" of the problem from the "optimization" of the decision .

Instead of a single model, they develop a menu of three distinct "information regimes." These regimes match the amount of data an agent can trust.

1. The Stationary Regime: This is the most structured benchmark. The agent assumes price adjustments follow a predictable Poisson arrival process (a model where events occur at a constant average rate). It also assumes new prices come from a known, stable distribution. The authors find the optimal strategy is a dynamic threshold policy. This threshold is a moving target. It is governed by an ordinary differential equation (ODE) that adjusts based on the time remaining until the deadline.

2. The Bayesian Learning Regime: Here, the agent knows how often prices change but is uncertain about the actual price distribution. As the agent observes new prices, it uses Bayesian updating (refining probabilities as new evidence emerges) to update its internal beliefs. The authors show the optimal rule remains threshold-based. However, the threshold now fluctuates based on the agent's evolving posterior beliefs (updated probabilities) .

3. The Robust Regime: This is the "safety first" mode for when data is sparse or markets are chaotic. The agent makes no probabilistic assumptions about the future. Instead, it relies only on known price bounds (the highest and lowest possible prices). It seeks to minimize worst-case losses. To prevent an adversary from exploiting a fixed strategy, the authors demonstrate that the optimal approach is a randomized threshold policy. The agent draws a random target price at the start of the window .

Performance in the wild

To test these mathematical rules, the researchers evaluated them against real-world Amazon price histories. Using data from the Keepa tracking service, they analyzed 367 items and 48,933 time-stamped price observations. They compared their specialized "OR" (operations research) policies against several baselines. These included simple "Buy Now" or "Buy Last" rules and standard Large Language Model (LLM) prompting.

The paper reports that the Stationary and Bayesian policies perform competitively in terms of mean normalized consumer surplus. This metric measures how much value the buyer captured relative to the best possible outcome. Interestingly, the Robust policy performed best at the 10th percentile of the surplus distribution . This means the robust model is superior at protecting against the worst-case financial outcomes.

One significant finding concerns the role of LLMs. The authors tested an "LLM-OR" agent. This agent uses a language model to select the best regime and calibrate the model. The mathematical policy then handles the actual execution. The results suggest LLMs are more effective at high-level reasoning. They excel at selecting which model to use and interpreting user requests. However, they are less effective at making granular "buy-or-wait" decisions directly.

Limits of the framework

The authors acknowledge several areas where the model remains stylized. The current formulation assumes a single-item purchase. This means it does not account for multi-item shopping carts or product substitution.

Real-world e-commerce also involves complex price movements. These might not perfectly mirror a Poisson process. Finally, the model does not currently incorporate stock-out risk (the possibility that an item becomes unavailable). For a practitioner building a production-grade agent, these omissions mean the system is currently a sophisticated decision engine rather than a complete end-to-end commercial solution.

The verdict: A hybrid architecture

If you are building an autonomous commerce system, the verdict is clear. Do not rely on an LLM to do the math. The research demonstrates that the most effective architecture is a hybrid one. Use the LLM for "soft" skills. It can interpret a user's natural language request and select the appropriate mathematical regime. Then, delegate the "hard" tactical decisions to the threshold-based OR policies.

The study provides a roadmap for moving toward strategic agents. You can deploy the Bayesian model to chase maximum savings. Alternatively, you can use the Robust model to ensure your agent protects against downside risk in a volatile market.

Figures from the paper