Original Summary

Hi HN. I partnered with Simon Kley to build this project at innovation intelligence hackathon (Berkeley).<p>Simply, we’re building a usefulness score on top of search for ai-agents.<p>Instead of agents reading 10 ranked relevant pages, we run a usefulness score first and give it only those that pass. We’re able to save token and cost with this and slightly better answers.<p>During our research, we found out that the current search results already do justice when it comes to simple queries. The rankings work well. They struggle when the task requires deep research or recency matters. So, we decided to talk to devs building ai-agents that research companies for sales outreach. Deep research and recency matters a lot to them. And in their words “every so often, stale facts reach emails.” They catch them by someone responding to the email or human review.<p>From our tests, When an agent got outdated alert for 7 companies, it pitched outdated news for 4 out of them.<p>Talking to the engineers already building this agents, we learned that cost saving is not the real goal. They’re ready to pay premium for confident results. So we added another check to our score. We call it superseding score, it check for pages replaced by newer news. For example, if two recent pages were fetched for news on Humane ai pin. If the first talks about how groundbreaking the product is and the second did that + talked about the HP acquisition, then the second page supersedes the first. The first page gets a lower score.<p>We’re looking for feedbacks here and also teams that are building agents that research companies for outreach. Thanks.<p>Plus, we rank a page once and every agent reuses it, so the more the agents use us, the cheaper the search gets.

  • 情报分类：技术学习与提效
  • 分类依据：内容涉及技术、AI、软件工具或工程实践
  • 信息来源：Hacker News 新项目
  • 发布时间：2026/10/13 12:07:07