Anthropic released Claude Code 2.1.326 on 16 November 2026. The headline is not a new feature. It is that the AI coding agent now stops and asks more often.
What the release notes say
- More approvals. Many Bash, PowerShell and file-write checks now ask for approval more often, or refuse outright.
- Credential files protected. Claude’s file tools refuse to write to sign-in files such as
.credentials.json. - Code Mode with agent teams. Code Mode, the experimental option where Claude writes JavaScript to call its tools, can now be turned on while agent teams are enabled.
- Visibility. A
/statusrow shows when a provider or an environment variable such asDISABLE_TELEMETRYhas switched features off. - Performance and fixes. Lower idle CPU use, faster typing and streaming in long sessions, Remote Control replies that stream to a phone or browser as they are written, and 233 bug fixes.
Why a non-developer should care
This follows 2.1.325 on 13 November, which also changed sandbox and git permission behaviour. The direction is consistent: vendors are making AI coding agents more cautious by default. For anyone commissioning software, that is reassuring, but it also means unattended automated runs may now pause waiting for a human.
So what
If a supplier says AI agents work on your codebase, ask how permissions are configured, who approves risky actions, and how secrets are kept away from the agent. Good answers are specific. Our AI-assisted development work is built around that kind of controlled workflow. To discuss it, start a project.