claude-code v2.1.323

Version 2.1.323 of Claude Code, Anthropic's coding agent, adds Code Mode, an experimental setting in which Claude writes JavaScript to call its tools. It stays off unless a user turns it on, and the release notes say an organization can also switch it off for everyone. Most of the rest of this long release is fixes, many of them to permission checks and sign-in handling, so Code Mode is the entry that changes how the agent goes about its work.

In ordinary tool use, each call is a separate step: Claude requests one action, such as reading a file or querying a connected service, the whole result comes back into its working context, and Claude reads it before choosing the next call. Anthropic set out the case for replacing that loop with code in an earlier engineering post about agents connected to many outside tools (our coverage of that post). A single script can make several calls, loop over their results, and keep only what matters, so the model sees what the code prints rather than every intermediate result, and an if-statement runs without waiting for the model to decide. The post also noted that Cloudflare had published similar findings under the name "Code Mode".

That post was about a general technique, and the release notes do not say how Claude Code runs the scripts, which tools they can reach, or how much context or time they save. The same post named the cost: running code an agent writes needs a sandbox, resource limits and monitoring that direct tool calls avoid. By its reasoning, the tasks that stand to gain are those with many calls or bulky results. Shipping the feature as experimental and opt-in, with an organization-wide off switch, means nobody's sessions change unless someone chooses to try it.