Every request a coding agent sends carries the same baggage: the system prompt and the definition of every tool the model may call. It is cached, so it is cheap to resend, but it still fills the context window and still counts. For Atlarix 15.11 we went looking for what nobody used and deleted it. The result is small, and that is the point of this post: it is about 4% of every turn, forever.
What went
The mobile tools. Atlarix had two tools for driving a phone emulator or simulator. They were rarely called, and a model can do the same with the commands a person would type (adb for an Android emulator, for example) through run_command, then look at a screenshot with view_image. So the special tools went, and the general ones stayed.
Image tools for models that can't see. Tools that return images are now offered only to models that accept images. A text-only model was being shown tools it could never use.
Onboarding. The first-run wizard, the guided tour, the coachmarks and the rotating tips are gone from the app, the CLI and the extension. The analytics choice still comes first. This part saves no tokens; it removes code and screens that few people finished.
Measured, not estimated
We counted the request a Build-mode turn sends before and after, in the CLI, with Claude Haiku 4.5 through Anthropic's format:
| Tokens per request | Tools offered | |
|---|---|---|
| Before (15.10) | 10,802 | 33 |
| After (15.11) | 10,350 | 31 |
That is 452 tokens and two tools fewer on every request, about 4.2%. A text-only model drops the image tools as well, so its saving is a little larger.
Why a small cut is worth it
Four percent sounds like rounding error, but it is paid on every model call, and a single agent message makes many of them as it works through tool results. Smaller models feel it twice: fewer tokens of instructions, and fewer tools to choose between, which is where weaker models make wrong calls. And every tool definition is a promise to keep working; deleting one we did not need is maintenance we stop paying for.
The rule we used
A feature stays if people use it or if the model needs it to do the job. If a general tool already covers the case, the special one goes. We checked usage first, then deleted, then measured on the real request, not on the source code, because what reaches the model is the only number that matters.
The cut shipped in Atlarix 15.11, alongside computer use and sessions that steer each other. A harness isn't what it carries; it's whether you can use it. The full list is in the changelog.
Originally published at atlarix.dev.
Top comments (0)