Claude Code 2.1.314, released on 30 October 2026, adds a setting called permissions.blockWritesOutsideWorkingDirectories. When it is on, Claude’s Write, Edit and NotebookEdit tools refuse any file path outside the working directories, in every permission mode.
Why the detail matters
Permission modes decide how often an AI coding agent asks before acting. Teams that let it run with fewer prompts, to move faster, have until now had less of a hard boundary around where it can write. This setting is a fixed edge: whatever mode is in use, file edits stay inside the project.
Other changes in the release
Idle conversations over 200K tokens are now summarised automatically in VS Code and the Desktop Code tab. FORCE_HYPERLINK=false now turns hyperlinks off, where it previously turned them on. Sessions that failed repeatedly with an “Invalid effort level” error, when an admin caps a model’s effort, now lower the effort and retry. And claude -p now prints an error and exits with code 1 on startup failure instead of exiting silently with code 0.
What it means
The pattern across recent AI coding releases is consistent: speed arrives first, then the controls that make it safe to use on real products. A boundary on where an agent can write is a basic control, and it is now one setting.
So what
If a supplier builds your software with AI agents, ask what limits they have set on what those agents can change. Guardrails like this belong in the working agreement, not in someone’s memory. Our AI-assisted development service defines them from day one. To start, start a project.