<?xml version="1.0" encoding="UTF-8"?><oembed><type>video</type><version>1.0</version><html>&lt;iframe src=&quot;https://www.loom.com/embed/a417f73c6bb84a1692985e12bfd6435f&quot; frameborder=&quot;0&quot; width=&quot;3024&quot; height=&quot;2268&quot; webkitallowfullscreen mozallowfullscreen allowfullscreen&gt;&lt;/iframe&gt;</html><height>2268</height><width>3024</width><provider_name>Loom</provider_name><provider_url>https://www.loom.com</provider_url><thumbnail_height>2268</thumbnail_height><thumbnail_width>3024</thumbnail_width><thumbnail_url>https://cdn.loom.com/sessions/thumbnails/a417f73c6bb84a1692985e12bfd6435f-0d3d8998796b546d.gif</thumbnail_url><duration>190.641</duration><title>Optimizing Search Routing for AI Agents</title><description>This Loom discusses routing search queries for AI agents across multiple search providers using feedback from live chat sessions instead of benchmark-driven auto-routing. The demo shows an optimizer initially selecting Brave, then submitting feedback after the user chooses to buy the snowboard. Adam explains that each query is classified into a search category using Jev and Google Trends, and providers are allocated using Thompson sampling based on historical performance to adapt over time while minimizing latency by moving computation off the hot loop. A simulation sends traffic at about 5 QPS and demonstrates how the system gradually shifts more query share to the best engine, eventually increasing usage of “Tivili.”</description></oembed>