Agents need contracts, not pages
A human can skim a result page and recover from messy formatting. An agent needs a stable contract: title, URL, snippet, score, source type, and enough metadata to decide the next action.
This is why a retrieval API should be boring at the transport layer. The interesting work belongs in ranking, freshness, and source extraction, not in unpredictable response shapes.
Scope is a product feature
A single generic web index is rarely enough. Product agents often need news for freshness, scholar for papers, document for PDFs, image for visual tasks, and video or podcast for media workflows.
Explicit scopes make the caller state intent. That gives the ranking layer a cleaner job and gives developers a better debugging surface when results feel wrong.
Token budgets matter
Every byte returned by retrieval competes with instructions, conversation history, tool outputs, and the final answer. Search APIs for agents should make concise snippets and raw content separate choices.
The default response should be useful enough for routing and citation, but small enough that an agent can call it repeatedly without exhausting context.