Giedrius Šteimantas uses a shopping-agent example to show why a capable model alone does not guarantee a reliable workflow. He divides the process into discovery, product verification and selection, a user decision, and purchase execution. Browser failures, expensive retries and unavailable products can prevent the system from completing the task even when its high-level reasoning is sound.
Giedrius Šteimantas recommends using a browser only where the interaction requires one. A fixed list of retailers can restrict discovery, while repeatedly loading heavy pages adds latency and makes transaction costs harder to predict. Product checks also need to reflect the customer’s location because stock, sizes and other options can vary across regions.
Giedrius Šteimantas emphasizes that an HTTP 200 response or a plausible content size does not prove a useful product page was retrieved. Feeding blocked or irrelevant pages into a language model wastes tokens and reduces the options available for a decision. He argues that content validity must be established before compression, rather than treating smaller output as a substitute for correct output.
Giedrius Šteimantas reserves browser interaction for the purchase stage, where dynamic pages and user inputs make it necessary. The agent presents its selection to the user first and proceeds only after approval. His broader lesson is to design the retrieval and interaction layers deliberately, so the model can work with valid information instead of absorbing infrastructure failures.
Watch on YouTube
