Cutting LLM tokens 10x for AI browser automation

How filtering the accessibility tree to ~50 relevant elements cuts per-action token usage from 30–50k down to 2–5k — with a benchmark against Stagehand on Amazon.de.