video1.0<iframe src="https://www.loom.com/embed/e621ff50ebb54ca48f9b8a1834b08d1e" frameborder="0" width="1720" height="1290" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>12901720Loomhttps://www.loom.com12901720https://cdn.loom.com/sessions/thumbnails/e621ff50ebb54ca48f9b8a1834b08d1e-e07aea69de694e54.gif147.093This Loom introduces Agent Lens, a system to stop AI coding agents from drifting away from front end and company requirements. The speaker explains that agents can build real software but often miss specific requirements, causing engineers to discover issues only after the agent finishes, when they must undo or redo work. Agent Lens uses a two agent loop with an executor that performs tasks and an evaluator that reviews results using a standing set of principles from prior corrections rather than a per task checklist. The team says they proved self correction by watching it fail four times with a weak model, and that Pydantic AI powers both agents and structured output with live tracing via Logfire.