TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents

Aaron J. Li¹, Weijia Xu², Chandan Singh², Zixiao Chen³, Jeevana Priya Inala², Charlotte Siska³,
Acar Tamersoy³, Bin Yu¹, Ion Stoica¹, Michel Galley², Jianfeng Gao²

¹UC Berkeley   ²Microsoft Research   ³Microsoft Security

Our measurement on SWE-Bench tasks: 18.4–30.9% lower mean cost across five agent–model combinations, cheaper in all 30 repository × combination cells, and Pass@1 up in 4 of 5. Results

TokenEconomizer makes coding agents cheaper to run in the repositories you work in often, without changing the agent itself. It learns from the sessions an agent has already run in a repository, and uses that experience to cut waste in the sessions that follow. It has three mechanisms:

• Micro-skills: remember how the repository works. Sessions keep rediscovering the same facts: a build step that must run first, the spelling of a command that works here (python -m pytest where pytest is missing, npx tsc where tsc is), an environment variable a command needs, where generated files live. Micro-skills learns such facts, in any language, from two sources: failures that repeat across sessions, and a model that reviews the recent sessions every k sessions (k = 3) and proposes facts. A fact is stated at the start of each session once three sessions have confirmed it, and dropped when later sessions contradict it.
• Tool pruning: stop loading tools that go unused. Every loaded tool's definition is sent with every request. A tool that goes unused over p prompts spread across at least m sessions, or over s sessions of any length (p = 10, m = 2, s = 5), is left out of the next sessions through the agent's own settings. One command brings it back.
• Failure-loop intervention (the loop guard): step in when a turn is stuck. When a turn (the agent's work on one prompt) keeps hitting the same failures, a judge model checks whether it is stuck. If the judge is at least c sure (c = 0.7), the agent gets a redirect (a short note, added to its context, to try another approach), at most r times per turn (r = 1). If the turn is still stuck after that, TokenEconomizer tells you how to stop it; it never stops a turn itself.

TokenEconomizer is a plugin for Claude Code, Codex, the GitHub Copilot CLI and Pi, with experimental support for opencode, Cursor Agent, Grok Build, Hermes Agent and the Antigravity CLI. Each italic letter above is a setting, shown with its default (Settings), and each mechanism can be turned on or off. The plugin announces each action in one line, and if any part of it fails, your session carries on as if it were not installed. Its model calls (the fact proposer, and the judge with the redirect it writes) run through your own agent, on the same plan or login, and their cost is included in its reports. What it learns stays on your machine, in ~/.tokeneconomizer.

Results

The paper ran five agent–model combinations on the same 277 SWE-Bench tasks from six repositories. Each task ran as its own session, and each repository's tasks ran in order, starting with nothing learned. Costs include every model call TokenEconomizer made. Within each repository and combination, cost fell by 5–40%. The savings held on Web-Bench (16–21% lower) and on two security-auditing benchmarks where every task is in a new repository (6–20% lower). The project page has interactive charts, per-repository numbers and nine annotated runs, and experiments/ reruns the SWE-Bench and Web-Bench experiments.

How it works

Coding agents are built to finish tasks, not to keep them cheap. Every turn pays again for the agent's instructions, every loaded tool, and the conversation so far. Developers use the same agent on the same repository again and again, so much of that cost repeats: the same facts rediscovered, the same unused tools sent, the same kinds of failing loops. TokenEconomizer leaves the agent's model, weights and code alone and learns across a repository's sessions instead of treating each one on its own, so whatever the plugin spends is paid back over the sessions that follow. Four principles guide the mechanisms:

• Work through what the agent already reads. Facts enter the context at session start, and tools are removed through the agent's own settings, so nothing depends on the agent learning a new tool.
• Count cost as it is billed. Most tokens in a sessio