ML//agent//hybrid agent

A hybrid agent is an AI agent that combines several sources and mechanisms (an existing search index, direct navigation, browser automation, APIs) to answer a question from the web. It does not traverse the Internet from scratch; at each step it picks the cheapest or most useful tool, so a typical run mixes search, reading, a browser and an API.

A hybrid agent is an AI agent that combines several sources and mechanisms (an existing search index, direct navigation, browser automation, APIs) to answer a question from the web. It does not traverse the Internet from scratch; at each step it picks the cheapest or most useful tool, so a typical run mixes search, reading, a browser and an API.

1A question2A search enginefive candidates3Reasoning4Two pages opened5One of them rendered in a browser6An API queriedone more figure7The answer

Each tool has a cost and a coverage: search is cheap and shallow, a fetch is cheap and precise, rendering is expensive and complete, an API is exact when it exists. The hybrid agent's skill is the routing rather than any single tool (tool use): choosing between search, fetch and render is choosing which observation to buy at which price, the general principle of active perception applied to the tools of the web.

The hybrid agent is the practical form of a web agent, and the reason the reading side of the web is no longer only crawlers and browsers but a third kind of client with judgment.