Claude Code has had a nagging usability problem since its debut some 13 months ago. Every file write and shell command requires explicit approval, which makes longer coding sessions slow and laden with interruption. Some developers work around this by disabling permissions entirely, but that opens the door to a wide gamut of unintended consequences.
Anthropic’s latest update is an attempt to bridge that gap, introducing a middle layer between constant approvals and full bypass.
Auto Mode for Claude Code: Permission checks on autopilot
The company is introducing what it calls “auto mode,” a setting that allows Claude to decide, case by case, whether an action should proceed. Instead of prompting the user for every step—or skipping checks altogether—the system evaluates each tool call before it runs.
At the center of this is a classifier that screens for potentially destructive behavior. Routine actions, such as small edits or standard commands, are allowed through automatically. More sensitive operations—like large deletions, suspicious scripts, or possible data exfiltration—are blocked. In those scenarios, Claude is expected to adjust its approach rather than push ahead.
If it keeps hitting dead ends, for example repeatedly attempting actions that are flagged as risky, the system escalates back to a human approval prompt.
Auto mode can be enabled from the command line using claude --enable-auto-mode, after which users can toggle into it directly from the terminal interface—shown in the animation above—by cycling through available modes with a keyboard shortcut.
It’s worth noting that auto mode builds on Claude Code’s existing permissions system, which allows developers to define rules for what should be allowed, denied, or flagged. The difference here is that the model is now handling most of those decisions on its own.
Anthropic describes this as a middle ground between the default mode, which asks for everything, and the --dangerously-skip-permissions flag, which asks for nothing.
The company says the system reduces risk but doesn’t remove it entirely. In cases where user intent is unclear, or where Claude lacks sufficient context about a given environment, potentially unsafe actions may still be approved. At the same time, some benign actions may be blocked.
Less interruption, with caveats
Auto mode’s appeal is fairly clear. Developers can run longer tasks without staying glued to the terminal approving each action. That opens the door to more autonomous coding sessions, where Claude can iterate, fix, and retry without constant oversight.
That said, the system still requires oversight for any meaningful production sessions. Anthropic recommends using auto mode in isolated environments, particularly for more sensitive codebases.
There are also trade-offs in performance. Each action now passes through an additional review step, which can add latency and increase token usage over time.
Auto mode: Another step toward autonomy
Auto mode arrives as Anthropic continues to push Claude Code toward longer-running, less supervised workflows. Earlier updates focused on multi-step reasoning and agent-style behavior, while this release focuses on how those actions are governed.
The underlying permission system hasn’t changed as such. Developers can still define rules, enforce denials, or bypass checks entirely. What changes is who handles those decisions in most cases.
Where developers previously approved actions one prompt at a time, auto mode delegates much of that responsibility to a classifier that checks each action before it runs and decides whether it should proceed.
The trade-off between safety and speed remains, but it’s now handled in the background rather than at the command line.
