Microsoft Research Plugin Cuts Coding Agent Costs Up to 31%
Microsoft Research said in a blog post on Nov. 16 that it has built TokenEconomizer, an open-source plugin that helps AI coding agents spend less. The work comes from UC Berkeley, Microsoft Research and Microsoft Security. In tests on 277 real bug-fixing tasks, it lowered average costs by 18.4% to 30.9%.
The idea is simple. Developers work in the same codebases again and again, but agents often start each session fresh. Every exchange with the model uses tokens, the units AI services use to measure and bill work. The plugin reads the log of each finished session and turns it into guidance for the next one. It does not change the agent or its model.
The researchers first studied Claude Code on 60 tasks from SWE-Bench, a set of real bug fixes in open-source Python projects. They found three kinds of waste. Once a repository had at least five finished tasks, 30% of the agent's tool calls repeated an earlier action there. The agent received 23 tool descriptions each turn, and 16 were never used. Tasks that failed were 17% of the total but 32% of the spending.
The plugin responds in three ways. It adds short notes about a repository once several sessions confirm a pattern. It removes tools the history suggests the agent won't need. And it checks whether an agent is stuck in a loop, then suggests a different approach. In the open-source version, if the agent stays stuck, the plugin alerts the developer instead of stopping the work.
The team tested five combinations of agents and models across six repositories. Costs fell in every repository, and success rates stayed broadly stable. Pass@1, a measure of tasks solved on the first try, rose in four of the five combinations. Early tasks can cost more while the plugin learns, but cumulative costs dropped below the default by the sixth task in every repository. Savings were smaller on other work: 16.0% to 21.2% on web development tasks and 6.4% to 19.6% on security auditing.
For people who use tools built on the list of AI models, the point is that cheaper agents could make this kind of help easier to afford. Microsoft Research did not say whether any product will use the plugin. It also did not say how the savings translate into dollars. The reported results come from benchmark tasks. How the plugin performs in everyday projects is still untested in the source. Microsoft Research said the plugin needs no API key of its own and keeps what it learns on the developer's machine.
Sources
Cite this
Microsoft Research Plugin Cuts Coding Agent Costs Up to 31%, New Mind Museum, November 16, 2026. https://newmindmuseum.com/research/2026/11/tokeneconomizer-coding-agents-cost/
New Mind Museum. (2026, November 16). Microsoft Research Plugin Cuts Coding Agent Costs Up to 31%. https://newmindmuseum.com/research/2026/11/tokeneconomizer-coding-agents-cost/
“Microsoft Research Plugin Cuts Coding Agent Costs Up to 31%.” New Mind Museum, 16 Nov. 2026, https://newmindmuseum.com/research/2026/11/tokeneconomizer-coding-agents-cost/.
@misc{new-mind-museum-2026-microsoft-research-plugin-cuts-coding-ag,
author = {{New Mind Museum}},
title = {Microsoft Research Plugin Cuts Coding Agent Costs Up to 31%},
year = {2026},
url = {https://newmindmuseum.com/research/2026/11/tokeneconomizer-coding-agents-cost/}
}Free to quote and cite with a link to this page.