Signal-Driven Observation cuts web agent context bloat by 10x

Long-horizon web agents traditionally ingest the entire DOM and accessibility tree — often exceeding 50,000 tokens — at every action step. This coupling of observation frequency to action frequency causes progressive context degradation, eroding reasoning before tasks complete. Shubham Gaur and Ian Lane's Signal-Driven Observation (SDO) challenges that architecture by introducing a dedicated sub-call that reads the full DOM once but returns only task-relevant elements and their selectors. A lightweight signal detector then determines when to re-invoke that sub-call, firing only on URL transitions, newly visible interactive elements, action failures, or exogenous browser events.

This approach draws inspiration from Recursive Language Models, where querying a document outperforms reading it wholesale. SDO decouples observation frequency from action frequency, drastically reducing token consumption while preserving agent accuracy. The paper outlines open problems, including optimal signal detector design and handling of dynamic content, and calls on the research community to treat observation compression as a core architectural decision in web agent design. For practitioners building autonomous web agents, this could slash costs and improve reliability on complex, multi-step tasks like form filling or data extraction.