Claude Code's new Code Mode is worth trying, and it changes what your permission rules and hooks actually see. If you run Claude Code on a team, here are five checks before anyone turns it on.
Code Mode shipped in Claude Code v2.1.323 as an experimental opt-in, off by default. Per Anthropic's Code Mode documentation, Claude writes a piece of JavaScript that makes several tool calls at once (search, edit, run tests) instead of asking for each call and reading each result. You see the tool calls in the conversation, not the code, unless you open the transcript viewer. It's faster on repetitive work. It also moves where your safety checks happen.
1. Decide who gets to turn it on. Each developer turns Code Mode on with the CLAUDE_CODE_EXPERIMENTAL_CODE_MODE environment variable. A project's settings file can't turn it on, so a cloned repo can't switch it on for you. An administrator can block it for everyone by setting the variable to 0 in managed settings, but can't force it on. If you want a pilot, not a free-for-all, keep it blocked in managed settings until you've picked the pilot group and set the rules below.
2. Check that your setup qualifies at all. The documented requirements rule out a lot of enterprise setups. Code Mode doesn't work through a gateway, Amazon Bedrock, or other cloud providers, or for organizations on zero data retention or the HIPAA configuration. It runs only in an interactive terminal session, not in -p, the Agent SDK, VS Code, the desktop app, or cloud sessions. If you route Claude through Bedrock for compliance reasons, the answer is no for now. Don't unwind a compliance decision to try an experiment.
3. Fix your hooks before they get compound commands. The docs give a clear example: in Code Mode, a git push inside a folder named app can reach your PreToolUse hook as cd 'app' && git push. A hook that matches the start of the command string misses that. Make your hooks check every part of a command. If you want to see the JavaScript itself, add a hook that matches REPL, the tool that runs the code, and read it from tool_input.code.
4. Know what auto mode reviews now. In auto mode, the classifier reviews each piece of code before it runs, not each tool call inside it. It doesn't see values the code computes at run time, such as a path taken from search output. So the calls you must control (pushes, deploys, anything touching production data) need deny rules or hooks, not trust in the classifier. We made the same case in why a classifier is not a sandbox, and it applies here twice over.
5. Pick a first task with a safe failure mode. The docs suggest tasks with many related tool calls: one change repeated across many files, or a search whose results need filtering. On a legacy codebase with no tests, that's also where a bad edit spreads fastest. Start read-only, like the docs' own example: list every Markdown link that points to a missing local file. Move to bulk edits on a branch, with a test run Claude can't skip, after you've read a few of the scripts it writes. That's the order we follow when we roll out any new agent capability on client code. Code runs for at most ten minutes unless it runs in the background, so long builds belong in a separate step anyway.
Code Mode is experimental, and Anthropic says it may change or be removed in any release. That's a reason to pilot it on purpose, not to ignore it. If you want help setting the rules first, our AI automation work starts with exactly these settings, and our checklist for proving an agent policy is enforced is the test to run after.

