web-browsing agent

A model's knowledge is frozen at the moment its training ended, so it cannot know today's news, this week's prices, or a page published yesterday. A web-browsing agent closes that gap by giving the model the ability to search the live web, open the resulting pages, read them, and follow links — gathering current information the same way a person researching a question would.

Under the hood it is a loop of tool calls: issue a search query, get back a list of results, choose and fetch a page, extract the relevant text, and decide whether to read more, search again, or answer. The fetched content becomes context the model reasons over, and good agents cite which pages a claim came from, so the answer can be checked against its sources.

The open web is messy and adversarial, which makes browsing risky: pages can be outdated, biased, or contain text deliberately written to mislead a reading model. So a browsing agent is only as trustworthy as its judgement about which sources to believe — retrieving a page is easy, knowing whether to trust it is not.