SecondSourceJudgment rebuilt from primary sources

According to the paper's abstract, author Abhishek Rai reports that open-weight agents that write scripts to call tools use 70% to 90% fewer tokens than agents calling MCP tools one at a time.

Concept illustration of a pull-down screen hanging from a black roller, with a black band across the top bearing a person's name in large white capitals. Below it, on a long white chest of drawers whose centre panel is lettered MCP, stand from left to right a laptop with a dark screen showing a beige speech bubble, a tray lettered TOKENS holding a tall stack and a short stack of thin horizontal layers, a laptop with a red screen of light horizontal lines, and a monitor lettered SCRIPT. Printed on the image: 70%–90% Fewer tokens: open-weight agents scripting tool calls vs one-by-one MCP calls, author Rai reports.

Beyond a thousand tools, the savings come from the script-writing approach as a whole, while an added routing layer that pre-filters tools gives no significant help, and only open-weight models were tested. The experiment uses 12 MCP servers and up to 1,958 distractor tools, testing agents from the Qwen, Gemma and Nemotron families. The author preregistered a hypothesis: first filtering a subset of tools with a cheap routing model, then layering on script-writing, would do better. The result was only a small, statistically insignificant improvement for weaker agents, and a slight cost increase for the strongest one (arXiv 2611.19123, 2026-11). How we checked →

Source and checks

This item ran in the brief of Nov 18, 2026, alongside 14 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.

Cite this item

SecondSource (2026). "According to the paper's abstract, author Abhishek Rai reports that open-weight agents that write scripts to call tools use 70% to 90% fewer tokens than agents calling MCP tools one at a time.." https://secondsource.io/en/research/2026-11-18-2/. Accessed access date.