Korva is a scroll of the internet’s best ideas.
One prompt, hundreds of searches
Searches per prompt
A typical search agent runs between five and twenty searches even when it answers within a few seconds. At the lower tiers of a chat app, every prompt triggers five to ten. Dial the setting up and it becomes hundreds or thousands. Just using an AI app multiplies a person's searching by about an order of magnitude.
Agents on a schedule
Bigger multipliers come from agents that run on a schedule. Take a lender with a portfolio of 10,000 small businesses and a monthly human process for judging whether each one's risk is rising or falling, which relies on web data. An agent can run that check weekly and produce a dashboard and a list of action items, and the number of searches goes up by hundreds of thousands or a million.
Meeting prep
The same pattern shows up on one person's calendar. A custom meeting-prep agent in Notion, built with a single prompt, searches the internal data and the web for every meeting, tens to hundreds of searches each time. Every new agent for a new use case multiplies the total again, and the only ceiling is whether the value is worth the spend.
Why does one prompt trigger several searches?
An agent works in a loop: the model plans, runs a search, reads the results, and decides whether it needs another. A broad question often breaks into several narrower ones that each need their own search. Reasoning models are trained to keep going until they judge the answer is complete. The number of searches therefore depends on how much the model finds it still needs to learn.
How do scheduled agents differ from chat?
A chat starts when a person types a message, while a scheduled agent starts when a clock says so. Scheduling on a timer is an old idea in computing: the Unix cron utility, which runs jobs at set times, dates to the late 1970s. An agent on a schedule needs no person present, so its work stacks up in the background. The only things that slow it down are cost and the time each run takes.
What limits how many searches an agent runs?
Each search result adds text to the model's context window, which has a fixed size. As the window fills, the model becomes slower and more expensive to run. Research from Stanford and others, published in 2023 as Lost in the Middle, found that models use information buried in long contexts less reliably. Agents therefore summarize or discard older results to make room.
More from Training Data
