Nidhi Vyas describes a commerce agent that treats vague requests as a collaboration problemAgentic commerce lets AI agents discover, recommend, and purchase goods or services within delegated financial controls. instead of jumping straight to product recommendations. The agent builds working state from session history, explicit constraints, inferred preferences and reference images, while tracking how confident it is in each assumption.
When important details are missing, the system chooses the next question by estimating information gainInformation gain measures how much an observation or answer reduces uncertainty about the decision an AI system is trying to make.. It can ask a focused text question or present a visual preference boardVisual preference elicitation learns what a user likes by asking them to react to images or visual choices instead of relying only on descriptions. when qualities such as style, fit or aesthetic are easier to recognize than describe. This is intended to uncover blockers without exhausting users with unnecessary questions.
The agent also changes its response format to match the taskResponse format selection chooses the presentation structure best suited to an AI task, such as concise text, a table, or a visual comparison.. A quick answer may use concise bullets, a comparison may need a table and a subjective choice may benefit from a visual board. Vyas says evaluation must therefore measure data fidelity, actionability and format selectionAgent evaluation measures whether an AI agent completes useful tasks correctly, safely, efficiently, and within its intended boundaries. rather than rely on one generic response score.
Vyas proposes tests for fact retention, confidence calibration, counterfactual sensitivity, blocker discovery, question utility and turn efficiency. She argues that stronger commerce agents will need shared merchant ontologies and agent-to-agent interfaces so they can exchange structured product knowledge while continuing to adapt to individual users.
Watch on YouTube
