One script instead of dozens of separate steps: Claude Code trials a mode in which the model drives its tools through JavaScript
Anthropic added an experimental Code Mode in version 2.1.323, and two days later it also changed how sandboxed commands ask for permission.
Anthropic has built an experimental Code Mode into Claude Code, according to the official changelog for version 2.1.323. With the mode switched on, Claude no longer calls each tool such as Bash, Read or Edit separately with a model response in between. Instead, Claude writes a JavaScript script that bundles those calls, and Claude Code runs it in the session. German site Clauding writes that the version appeared on npm on the evening of 12 November and gives the example of reading twenty files and running three commands: normally 23 separate steps, with a script a single pass. That saves tokens and waiting time.
Developer Daniel Ávila describes on DEV how it works. The mode is enabled with the environment variable CLAUDE_CODE_EXPERIMENTAL_CODE_MODE in the user's own settings; it is ignored in project files, so a cloned repository cannot turn it on. According to Ávila it currently only works in the terminal, with recent models such as Opus 5.5, Sonnet 5.5 and Fable 5.2, and not through the Agent SDK, the desktop app, cloud sessions or Bedrock and Google Cloud. Deny rules and hooks still check every call, but in auto mode the classifier assesses the script before it runs rather than each individual call. Clauding points out that a script is harder to review than a list of separate steps.
Adjustments have already followed. Version 2.1.325 makes sandboxed Bash commands follow the normal permission flow by default again; anyone who wants them to run without asking now has to configure that. Version 2.1.326 also allows Code Mode when agent teams are enabled. Anthropic stresses that it is an experiment that may change or disappear in any release.
Transparency: this edition was compiled by Claude, made by Anthropic.