Thursday, November 19, 2026 Updated Briefings Saved 中文
AI Daily Radar

AI and the frontier tech around it, explained every morning

Released

TokenEconomizer Cuts Coding-Agent Costs by Learning From Past Sessions

Photo: Innovalabs / Pixabay

Microsoft Research, UC Berkeley, and Microsoft Security have introduced TokenEconomizer, an open-source plugin that learns from an agent’s completed sessions in a repository and adjusts future sessions without changing the agent or its underlying model. It uses one-line repository notes called micro-skills, removes tools that history suggests are unused while retaining core command and file-editing tools, and intervenes when a judge model decides an agent is stuck in a failure loop. In the authors’ evaluation across 277 real bug-fixing tasks and multiple coding agents, it reduced costs by 18.4–30.9%, with costs falling in every repository tested and task success rates remaining broadly stable. The supplied post does not name the specific agents or models tested or provide installation and general-availability details.

Verified and unconfirmed

  • Verified: Microsoft Research’s post states that TokenEconomizer is an open-source plugin that learns from session logs without modifying the agent or model, and that in the authors’ 277-task evaluation it reduced costs by 18.4–30.9% across multiple coding agents, with costs falling in every repository and success rates broadly stable.
  • Not yet confirmed: Independent reproduction, the specific agents and models tested, absolute cost or token figures, and public installation or availability details; the supplied excerpt is truncated and does not provide them.

Background

Coding agents such as Claude Code resend instructions, tool descriptions, and session history to the model on every turn, so repeated work in the same repository adds token cost. The authors profiled Claude Code on 60 SWE-Bench tasks—real bug fixes in open-source Python projects—and found that after a repository had at least five completed tasks, 30% of tool calls repeated an earlier action in that repository, 16 of 23 supplied tool descriptions were never called, and the 17% of tasks that failed consumed 32% of spending. TokenEconomizer targets that repository-specific waste rather than a single task’s prompt.

Why it matters

For teams running coding agents repeatedly in the same codebase, TokenEconomizer offers a way to cut token spend without switching models; the reported 18.4–30.9% reduction would matter most on high-volume bug-fix work, but teams should verify that the broadly stable success rate holds for their agent and repository before relying on the savings.

Comments

Comments

Comments use a GitHub account and are moderated. Comment rules

More in AI

See all