- Author
- Dimitri Mazmanov, Principal Product Manager, Spotify
- Published
- 3 September 2026
- Kind
- Company engineering blog post with a how-to and a small benchmark
- Original
- engineering.atspotify.com/2026/9/portal-by-spotify-cut-my-claude-code-token-usage-by-90
- Subject
- Harnesses and agent loops
- Themes
- Small means
What it is
Dimitri Mazmanov, a principal product manager at Spotify, wrote this how-to. It shows two Portal agent modes and a plugin named shunt. It aims to cut Claude Code token use by routing grunt work.
What it says
- Mazmanov says much of a coding agent's work is I/O, not reasoning.
- The post routes bulk reading and boilerplate writing to a cheaper model.
- A plugin called shunt uses hooks to block reads of large files and redirect them.
- Tests on a Java monorepo showed mean bulk-read savings near 90%, the post says.
- The post says delegation does not work for editing or for reasoning tasks.
- Per the post, responses typically take 10 to 30 seconds, so small reads stay local.
- The post says the bulk-reader and code-writer modes are public and shareable.
What evidence it gives
The post gives a small benchmark on one Java monorepo across four scenarios. It reports the mean saving but not per-scenario figures. Code-writing savings are said to be hard to measure in tokens. It shows configuration and script examples and one anecdote of a missed bug.
Interests and limits
Spotify publishes this about its own Portal platform and AiKA modes. The author is a product manager there. The savings count Claude tokens, not total cost or output quality. The setup needs a Portal instance, so outside readers may not reproduce it.