The agent had been caching its tools and system prompt, but it removed screenshots and trimmed older text on each step. Those edits changed the request prefix and prevented reuse of the growing history.
The team cached the history with a marker on the latest tool result, used batch screenshot pruning at a 20:1 ratio, and raised the history budget from 120,000 to 480,000 characters. The study tested six caching and history policies at two budgets on GPT-6.1 Sol and three unnamed frontier models. It comprised 144 runs, with three runs per condition, plus a 12-run follow-up.
StackAI reports that its best configuration was 29 times cheaper and four times faster than the production baseline on Model B. On GPT-6.1 Sol, it reports a 76-fold cost reduction, a fivefold speed increase, and 89% of input read from cache. The authors say every best-condition run found all 192 facts.
The source does not establish that these results transfer to other browser tasks, providers, or longer-running agents. Three tested models are unnamed, and the authors note that the run counts do not distinguish conditions separated by only a few percentage points.