Conversion Labs 用 Jev 做工具检索，从 100+ 工具里选对调用

Neil Tewari @NeilTewari原推发布于 2026.09.30

在 X 上打开原推

Today we're announcing @tryconversion Labs, our new AI research initiative, alongside our team's results in optimizing context for long-running agents. Our small team of AI engineers took roughly 70,000 tokens out of every request our agent makes, which cuts about two seconds off every turn. They've been researching how to optimize agent context, which models best tackle long-horizon agent tasks for marketing use cases, and how to orchestrate agents amongst complex datasets with 100+ potential tool calls triggered by ambiguous queries. Building an audience from a plain-English request used to take frontier models like Claude 45 seconds. Our agent now does it in about 2, with the same accuracy, and it's already live for our customers. If you're curious, our team @vikramsiva29, Leo Tian, @levibkline ,and @jimfutsu has produced two releases: (1) One-to-One: our benchmark for how well AI personalizes marketing content (2) Tool search with Jev: how we pick the right tools for our agent out of 100+ We're especially proud of these releases because the team worked deeply alongside our product team to define how agentic experiences actually touch the lives of our customers. Marketing is one of the hardest use cases for agents as success criteria are qualitative and no two companies model their data the same way. See more at https://conversion.ai/labs/. If you're an engineer curious about how technology touches the lives of real enterprise marketing and growth teams, we're hiring.