What do teams build with a web search API for AI agents?

What do these four jobs have in common?

Each one needs the words on the page, not just a list of links. A snippet of 160 characters rarely holds the figure, the date or the condition a model must quote. So most stacks end up with two services: a search API to find URLs and a reader to fetch them.

LLMWebSearch merges the two steps. One call runs a live web search and reads the first pages it finds. Each result carries its position, title, URL, host, snippet and, when it was read, the page's main content as Markdown or plain text with its word count and publication date when known. The agent gets text it can cite in the same round trip, and you get one bill.

The other shared need is control. You decide how many results come back (max_results), how many pages are read (read) and how long each page may be (max_chars). Those three numbers set both the token cost on your model's side and the credit cost on ours.

Which settings fit each use case?

Start from the defaults and change one thing at a time.

• Agent tool. Keep max_results=5 and read=3, and cap pages at 3,000 to 5,000 characters. Agents call the tool often, so short answers keep the context clean.
• Deep research. Many small searches beat a few large ones. Use read=3 per query and run more queries. Raise read only when one query must cover a narrow subject in depth.
• RAG. Ask for longer pages (max_chars around 12,000) and chunk them yourself. Add include_domains when answers must come from known sources.
• News monitoring. Set topic=news and time_range=day, then remove URLs you have already seen. Read three articles per query to summarise what changed.

All filters are listed in the search reference.

How do I size a first pack?

Multiply tasks per day by searches per task, then by 30. An assistant that answers 100 questions a day with 4 searches each needs about 12,000 searches a month. That fits in 25,000 searches with room for a second month.

Credits do not run out at the end of a billing cycle. Search credits have no expiry date: what you do not use this month is still there next year. A small pack bought for a prototype keeps its value until the project grows. The cost guide works through three larger cases and compares them with other APIs and with the search tools built into models.