<?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/8def4bf3ebf94a31a61b49b9fdad3699&quot; frameborder=&quot;0&quot; width=&quot;1916&quot; height=&quot;1437&quot; webkitallowfullscreen mozallowfullscreen allowfullscreen&gt;&lt;/iframe&gt;</html><height>1437</height><width>1916</width><provider_name>Loom</provider_name><provider_url>https://www.loom.com</provider_url><thumbnail_height>1437</thumbnail_height><thumbnail_width>1916</thumbnail_width><thumbnail_url>https://cdn.loom.com/sessions/thumbnails/8def4bf3ebf94a31a61b49b9fdad3699-d0294c944f7bd59c.gif</thumbnail_url><duration>166.741</duration><title>TokenWise Agents Self Optimize With Tracing</title><description>This Loom explains TokenWise’s self-optimizing agent system that uses memory and full tracing to avoid repeating mistakes. The agent runs in three loops at three speeds, catches issues in seconds, escalates in minutes, and learns quietly in the background while checking cached, unredacted trace payloads. It stops unsafe shell commands using SyngrabGuardian, retrieves teammate lessons from team memory, and locks fixes in storage; repeat errors trigger logging into ClickHouse and creation of new lessons into Sanskrit, with version history preserved. A brand new session can then apply yesterday’s fix on the first push, and checks turn green after three failures before one push after, with over 500 tool calls and 500 security scans traced so far.</description></oembed>