AI Aulis

AI NEWS IN ENGLISH · 17 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 17 November 2026Paper

TokenEconomizer cut coding agents’ costs by 18.4–30.9 per cent in its authors’ SWE-Bench test

AI-generated contentWritten by AI

msft.it, x.com · 2 sources · Sources published: 16 November 2026

An open-source add-on from Microsoft Research and UC Berkeley learns from a coding agent’s earlier sessions in the same repository and cuts unnecessary work. In the authors’ tests, costs fell and task success stayed largely unchanged.

Microsoft Research presented the TokenEconomizer add-on on its blog and X account on 16 November. It developed the add-on with UC Berkeley and Microsoft Security. The add-on runs alongside a coding agent and does not modify the agent or its model.

The researchers’ starting point is that an agent often begins a session from scratch, even though a developer works in the same repository day after day. They ran Claude Code on 60 tasks from the SWE-Bench benchmark, which consists of real bug fixes from open-source Python projects. Once at least five tasks had been completed in a repository, 30 per cent of tool calls repeated an action from an earlier task. On every turn the agent received descriptions of 23 tools, 16 of which it never called. Failed tasks made up 17 per cent of the total but accounted for 32 per cent of costs.

Three ways to cut waste

After each session, TokenEconomizer reads the session log and updates what it knows about the repository. It uses three methods:

  • One-line notes called micro-skills describe the repository’s quirks. The add-on adds a note to the agent’s initial context only once several sessions have confirmed the observation. If later sessions contradict it, the add-on removes the note.
  • Tool pruning removes tools that, judging by earlier sessions, the agent is unlikely to use. Basic tools for running commands and editing files stay, and removed tools can be restored.
  • When the agent repeatedly hits the same error, a judge model decides whether it is stuck. If it is, the agent gets a short suggestion to try a different approach. If it stays stuck, the open-source version notifies the developer rather than stopping the work itself.

Costs fell, success stayed largely unchanged

The authors tested the add-on on 277 SWE-Bench tasks from six repositories with five agent–model combinations. In each repository, the test started with no session history. Mean cost per task fell by 18.4–30.9 per cent compared with default settings, and the figures include the cost of the add-on’s own model calls. The results are the authors’ own.

The share of tasks solved on the first attempt (Pass@1) rose in four of the five combinations. Success rates stayed largely unchanged, the authors said. Lowering reasoning effort saved as much or more but markedly worsened results.

The savings are not immediate: early tasks can cost more while the add-on is still learning. With Claude Code, cumulative costs dropped below the default in every repository by the sixth task at the latest. On Web-Bench web development tasks, costs fell by 16.0–21.2 per cent. In security audits, where each task was in a new repository, the reduction was 6.4–19.6 per cent.

The add-on is available as open source on GitHub. It needs no API key of its own and keeps what it learns on the developer’s machine.

Report an error in this story