Formulate a simple instruction to the web agent. Return ONLY JSON: {"goal":"what you need to get in the end","why_agent":"why is an agent needed here, and not a regular search","expected_result":"what the finished result looks like, which shows that the problem has been solved"}
A regular search returns a list of links - then you open, read, compare. The web agent is designed differently: under the hood is a language model that reads the page like a human, plans steps, clicks buttons, follows links, and returns a ready-made response. Imagine an intern: you say “go to three stores, find out where delivery is cheaper,” and he brings the result, not ten tabs. The mechanics are simple: the model sees the contents of the page, decides what to click next, and repeats the “look-think-do” cycle until it completes the goal. Working technique: always set a “sign of readiness” - by what specific result will the agent understand that the job is done. Insider: The agent is more accurate when the goal is measurable (“three options with a price and a link”) rather than vague (“something good”). A typical rookie mistake is to write “best laptop” as a search engine query. The agent needs an order with a specific outcome, not a set of keywords.