THE ESSENTIALS IN 30 SECONDS

  • Background subagents in Claude Code show a status line refreshed about every 30 seconds; Haiku 5.5 now writes it from compact notes instead of re-reading the full conversation.
  • Anthropic estimates each update is about 99% cheaper and that sessions use about 5% less of their usage limit, or roughly 6% more prompts.
  • Reports say the change needs Claude Code v2.1.316 or later; foreground subagents are not affected.

Anthropic has changed how Claude Code reports on subagents that run in the background, a move it says reduces how much of a user's usage limit those agents consume. The announcement came on November 4 from the company's ClaudeDevs account on X and was relayed by several outlets, including Gizmodo Japan, Latent Space and AlphaSignal.

What changed

When Claude Code hands a task to a background subagent, the interface shows a one-line status that refreshes about every 30 seconds. According to the reports, each refresh previously sent the subagent's entire conversation to be summarised, so the cost of a single update grew as the transcript lengthened. Under the new approach, the subagent keeps short running notes and Haiku 5.5, a smaller and cheaper Claude model, reads those notes to write the status line. The model doing the subagent's actual work is unchanged.

Anthropic's figures, as reported, are that each status update now costs about 99% less, and that sessions using background subagents draw roughly 5% less of their usage allowance, which the company equates to about 6% more prompts within a limit window. Latent Space and AlphaSignal say the change requires Claude Code version 2.1.316 or later and needs no configuration. Gizmodo Japan notes that only background subagents are concerned: foreground subagents, which the main agent waits on, do not get these periodic summaries.

What it tells users

The announcement also shows that a display feature had a measurable cost. If refreshing a status line accounted for around 5% of usage in affected sessions, people who run many agents in parallel were spending part of their quota on progress messages rather than on the tasks themselves. AlphaSignal adds that the optimisation works independently of the CLAUDE_CODE_SUBAGENT_MODEL setting, which chooses the model subagents use for their tasks.

What is not confirmed

The ClaudeDevs posts on X could not be opened directly for this article, so the figures above rely on the outlets that quoted them. Claude Code's public changelog lists several recent subagent changes, including an entry in version 2.1.318 about how subagent progress summaries behave on Bedrock, Vertex and Foundry or with telemetry disabled, but it does not spell out the switch to Haiku 5.5 or the savings estimates. The 5% figure is an average set by Anthropic; actual savings will depend on how many background subagents a session runs and for how long.