Search Tool Adds Usefulness Scoring for AI Agents

Summary

A project built by two developers at a Berkeley innovation intelligence hackathon adds a usefulness score to search for AI agents. Instead of giving agents a conventional list of relevant pages, it filters pages first and passes on only those that meet the score, which the builders say reduces token and search costs while producing slightly better answers. Their research found that current rankings work well for simple queries but struggle with deep research and time-sensitive tasks. Engineers building agents for company research and sales outreach told them that stale facts sometimes reach emails and are discovered only through replies or human review. In one test, an outdated alert affecting seven companies led an agent to pitch outdated news for four of them. The project therefore adds a “superseding score” that lowers the ranking of an older page when a newer page contains its information plus a material update, such as an acquisition. The builders say users value confident results more than cost savings alone. They also rank each page once and let multiple agents reuse the result, potentially lowering search costs as usage grows. They are seeking feedback and teams working on company-research agents.