The paper introduces CAVEAT, a for in nine marketplace environments with eight common steering mechanisms such as sponsored ranking and scarcity cues. Without steering, agents purchase the user-optimal product 78.6% of the time; with steering, that drops to 17.3%. Even the most capable models show large degradations. Trajectory analysis reveals three failure modes: premature search closure, objective drift, and premature commitment. This work identifies incentive robustness as a distinct challenge for delegated agents, quantifies the gap, diagnoses specific failures, and demonstrates that inference-time harness and targeted post-training can substantially improve robustness.
Marketplace Steering Often Diverts Computer Agents From Users’ Best Choices
Across controlled shopping environments, common promotional mechanisms sharply reduced agents’ selection of products that best matched user requirements.
Big Tech
Yuxuan Li · Will Epperson · Wesley Deng · Zezhou Huang
Carnegie Mellon University · Microsoft Research
Research Digest··2 min read
In one line
Computer-use agents often fail to purchase the user-optimal product when online marketplaces use steering mechanisms.
Before you read this
Three ideas this paper builds oncomputer-use agentAn AI agent that interacts with computer interfaces—GUI, terminal, or code—to autonomously perform tasks.50 papers →benchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.366 papers →agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.186 papers →
Read it as
Li and colleagues introduce CAVEAT, a benchmark testing whether computer-use agents preserve user objectives when online marketplaces favor competing outcomes.
The authors built nine controlled marketplace environments in which agents search, compare and purchase products.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.