Skip to main content
Answer a broad question from several angles without filling your context. Your code splits the question into focused searches and runs them at once, merges the results so each URL appears once, and reads only the top pages, each cut at 3,000 characters. Run it again with follow-up searches when the pages leave a gap.

Recipe prompt

Paste this into your coding agent to have it build the recipe in your project:
Recipe prompt

What you build

When it runs

Once per research step. A step costs one search per sub-question and one fetch per page you read: 3 searches and 3 fetches in this sample. Searches and fetches share one rate limit, so the step size decides how fast you can loop.

Code

This sample runs three searches at once, merges them and reads the top 3 pages. It runs with no API key:
The TypeScript sample needs keenable 0.2.0 or later on npm, for the maxChars option on fetch. Python uses AsyncKeenable with asyncio.gather, TypeScript Promise.all, Go goroutines and Rust tokio::spawn on a cloned client, which shares one connection pool. The curl tab runs the searches in the background and merges their URLs with awk, so its merged order follows whichever search answers first.

Example output

You see how many unique pages the searches found, then one line per page read, with its length after the cap:
Example output · results vary
To hand the pages to your model with numbered citations, render each one with the page’s to_context(), or use search results’ to_context() when snippets are enough. See Context size.

Production notes

Next steps