AI - Claude Code background subagents get 99% cheaper status updates and Claude desktop gets /storage

Two small but practical Anthropic updates this week 🤣 Claude Code has made background subagents cheaper to run, and the Claude desktop app can now show you how much disk space it is quietly eating.

Neither is a headline feature, but those are often the changes that make a real difference day to day. The first one made me smile in particular, as it turns out that a surprising amount of usage was being spent on a subagent telling you what it was doing 😊

So far /storage does not work in VSCode extension, only in the desktop app 😭 I hope it will come to the extension as well in the future 😊

Claude Code: cheaper background subagents

Official announcement

Anthropic’s Claude Developers account announced the change on X:

Background subagents in Claude Code now use less of your usage limit.

A running subagent updates its status line every 30 seconds, and each update used to re-read its whole conversation. Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less.

The graphic attached to the post adds an important detail about scope and the overall saving:

Claude Code CLI. Across a whole session, about 5% less of your limit.

What is new?

• Background subagents refresh their status line every 30 seconds. A subagent that runs for an hour would therefore trigger around 120 updates.
• Each update used to re-read the subagent’s whole conversation. The longer the subagent ran, the more it had to read every single time.
• Haiku 5.5 now writes the status line from short notes instead. Anthropic says each update costs about 99% less.
• Across a whole session, Anthropic says the saving is about 5% of your usage limit. The graphic labels this as the Claude Code CLI.

Claude Code’s documentation confirms that each subagent sends its own requests, and that those count towards the same usage limits as your main conversation. That is why this matters: a subagent was spending your limit on housekeeping, not only on the actual work.

My take

The two numbers are easy to mix up, so it is worth separating them: 99% less per status update, but only about 5% less across a whole session. Both can be true, as the status line was only one part of what a session spends. A 5% saving is still welcome for something you did not have to change anything to get 😉 Some posts convert it into “roughly 6% more prompts”, but that is a third-party estimate. Anthropic’s own graphic only says about 5% less of your limit.

I also like the approach. A small, cheap model handles a small, cheap job, and the big model does not have to re-read an enormous conversation just to write one line. I covered Haiku 5.5 and Anthropic’s suggestion that it pairs well as a subagent, so it is nice to see it used for exactly this sort of housekeeping. It follows a similar theme to the idle compaction change I mentioned in my post on Claude Code desktop and parallel sessions: stop paying to re-process context that you do not need to re-process.

Of course, this only helps if you actually run background subagents. If you do not, there is nothing to see here 🤣

Relevant X posts

Update advice and summaries

Mirza’s advice is the practical one: update now, as background subagents were re-reading the entire conversation just to refresh a status line. His post also mentions the 5% figure:

Pranto C. gives a quick summary of the numbers, and also suggests a minimum Claude Code version that I could not verify against the public changelog:

Timurs Jeparskis makes the important distinction between the saving per update and the overall saving:

But here’s the important distinction: that doesn’t mean Claude Code uses 99% fewer tokens overall.

Freddy Montes summarises the change in Spanish:

CLI only, and a similar idea for your own agents

Anthony Kroeger notes that this looks like a Claude Code CLI-only problem and jokes that it is another win for team GUI:

Paweł Huryn says Anthropic did for the status line what he does for the agent itself, by keeping state in a small file instead of re-reading the transcript:

Same idea really. Don’t re-derive state from the transcript. Keep it in a small file.

Claude desktop: /storage

Official announcement

Lydia Hallie, from the Claude Code team at Anthropic, announced the new command on X:

We recently added /storage to the desktop app!

You can see how much disk space Claude is using on your computer and clear out inactive tasks and unused worktrees.

What is new?

• A new /storage command in the Claude desktop app shows how much disk space Claude is using on your computer.
• You can clear out inactive tasks and unused worktrees from the same screen.
• According to Alex, the usage is split by worktrees, conversation history, caches and Claude Code versions. That breakdown comes from his post rather than from Anthropic.

I could not find /storage in Claude’s desktop documentation at the time of writing, so the details above come from the posts below.

My take

This feels like a natural follow-up to the desktop app’s focus on parallel sessions. When each session can work in its own Git worktree, as I covered in the parallel sessions post, those worktrees can add up quickly, and it is easy to forget they are sitting on your disk 🤣 Being able to see where the space is going and clear it from the app itself is a sensible addition.

One bit of practical advice from me: before clearing any worktrees, check that nothing you still need is sitting unmerged in them 😉

Relevant X posts

TestingCatalog also covered the new command:

Alex lists the breakdown by worktrees, conversation history, caches and Claude Code versions:

Useful links

• Manage costs effectively (Claude Code documentation)
• Create custom subagents
• Claude Code desktop documentation
• Claude Code changelog on GitHub
• Claude Code release notes
• Introducing Claude Haiku 5.5