U.com CTO argues agent search needs distinct infrastructure from human search

Sahil Jain, CTO of U.com, argues that search infrastructure for AI agents must be fundamentally distinct from systems designed for human users. He highlights that agents can execute 1,000 to 10,000 searches simultaneously, generating high queries per second levels that human users do not match. While humans typically consume only the first few results and rely heavily on ranking order, agents prioritize recall; if information appears in the top 10 results, the agent can utilize it regardless of its specific position. Jain noted that recall generally seems to matter more than precision when it comes to agent results.

U.com, which began as a consumer search engine in 2020, has pivoted to building independent search infrastructure for agents. The company recently released a 'knowledge' layer that supplements public web search with authoritative, non-indexed data from licensed sources. To address the limitations of static benchmarks like BrowseComp, which allow for answer memorization, U.com is developing dynamic benchmarks such as LiveQA and VerticalRTK. These new tests require fresh data procured hours later to ensure accuracy. Additionally, the company post-trained NVIDIA's Nemotron Lightning model using their search tool to improve performance on agentic benchmarks. Jain maintains that search remains a fundamental tool for AI models, as they require external information to make better decisions and overcome training data cutoffs.