A new environment variable, CLAUDE_CODE_SUBAGENT_EFFORT, lets you set a lower effort level for subagents independently of the parent session. Set it via the env var or customize it in your agent config.
A software factory pairs a high-level orchestrating skill (like implement-ticket or fix-bug) with an automatic trigger that fires on new tickets or log errors, rather than waiting on a human prompt. Adding verification on top of that is what makes it actually work.
An Anthropic team member (adocomplete) enthusiastically co-signs the return of the ultrathink keyword to the Claude Code CLI, which had been temporarily disabled.
After reviewing session IDs submitted by users, the Claude Code team found no single root cause or regression for faster-than-expected token consumption. Most heavy usage traced back to one or two sessions using high effort, /goal, Workflows, Agent Teams, or long SDK jobs. A few smaller bugs that caused some sessions to burn tokens faster were found and fixes are rolling out. The team advises watching your effort level since subagents inherit it by default, and suggests using /usage to audit token spend.
Verification matters for any agentic use, but it becomes especially critical in factory and loop engineering, where automated pipelines can compound errors without a human in the prompt chain.
Lydia Hallie reacts warmly to a video related to the ultrathink CLI keyword returning to Claude Code.
When Claude creates a document in the desktop app, it tells the user to look above for the document even though it appears below. A Claude team member acknowledged the bug.
