DOM grounding
DOM grounding is the step where an agent's intended action, expressed in language or coordinates, gets mapped to an actual element in the page: "click the submit button" becomes a specific node. Grounding failures are a dominant cause of agent errors, and recorded DOM state is the ground truth they are diagnosed against.
How it works
Between the model's intent and the browser's execution sits a resolution problem: the page may contain three buttons that say submit, the visible one may be covered by a modal, or the element may have re-rendered with a new identity since the model observed it. Grounding strategies vary (accessibility roles and names, generated selectors, set-of-marks numbering on screenshots, coordinates), and all of them fail in characteristic ways that only the interface state at action time can explain. This is where recording earns its place: an rrweb stream captures the DOM as it was at the grounding moment, so a failed run answers "what did the page contain when the agent resolved that reference" directly. Comparing grounded targets across runs of the same task also exposes flaky grounding before it becomes a production incident.
Where you encounter it
Agent debugging sessions, framework design discussions, and eval failure taxonomies, where grounding errors get their own category.