Agent observability
Agent observability is the practice of capturing what an AI agent did and why, in enough detail to debug, evaluate, and audit its behavior. For agents that operate software through a browser, it spans three layers: the model's reasoning and tool calls, the actions taken, and the resulting state of the interface.
How it works
Conventional observability covers the first layer well: LLM traces record prompts, completions, and tool invocations. The gap is the third layer, because "the agent clicked submit" in a trace says nothing about what the page actually showed, whether a validation error appeared, or what state the form was in. DOM-level session recording closes that gap: an rrweb stream of the agent's browser session captures the interface's ground truth as structured data, aligned by timestamp with the trace. The combination answers the debugging question traces alone cannot: the model's action was reasonable given what it saw, or was not, and here is what it saw.
Where you encounter it
Agent development and operations, evaluation pipelines, and incident review; the broader context is covered in the session replay infrastructure for AI agents article.