Anthropic’s Claude Code preps Auto Mode to cut prompts

With the launch of Auto Mode, Anthropic is looking to help Claude Code run longer tasks without constant permission pop-ups. The research preview is slated no earlier than March 11, 2026, adding prompt-injection safeguards and a clearer admin off switch.

March 11, 2026

•

Claude

TL;DR

Claude Code “auto mode”: Automatically handles permission decisions to reduce repeated approval prompts during long runs
Release timing: Research preview, no earlier than March 11, 2026
Positioning: Middle-ground between manual approvals and --dangerously-skip-permissions
Security: Additional safeguards against prompt injections during tool-using sessions
Trade-offs: Not foolproof; recommended only in isolated environments; slightly higher token usage, cost, and latency
Controls: Enable via claude --enable-auto-mode; admins can block with "disableAutoMode": "disable" via Claude.ai, MDM, or managed JSON files

Anthropic is preparing a new permission setting for Claude Code, aimed at a familiar pain point in AI-assisted coding: long-running tasks that keep getting interrupted by approval prompts. The feature, called auto mode, is slated to launch in a research preview no earlier than March 11, 2026, and is positioned as a middle ground between strict manual approvals and the more blunt --dangerously-skip-permissions approach.

What “auto mode” changes in Claude Code

In auto mode, Claude can handle permission decisions automatically during coding sessions, reducing the need for developers to constantly confirm actions during extended runs. Anthropic also notes that auto mode comes with additional safeguards against prompt injections, reflecting a broader concern for tool-using models that can be steered by untrusted content encountered during a session.

That said, Anthropic is clear about the trade-offs:

Auto mode isn’t perfect and won’t catch every risky action
It’s recommended only in isolated environments
It can slightly increase token usage, cost, and latency

The overall framing is pragmatic: auto mode is meant to be safer than bypassing permissions entirely, without claiming to eliminate risk.

Enabling auto mode (and what admins can do about it)

By default, users will be able to turn on auto mode by running claude --enable-auto-mode, with no other action required to make it available.

For organizations that prefer to block it, Anthropic outlines multiple ways to disallow auto mode by setting:

"disableAutoMode": "disable"

Server-managed settings (Claude.ai)

Admins can go to Admin Settings → Claude Code → Managed Settings within Claude.ai and add the key/value configuration:

"disableAutoMode": "disable"

MDM/OS-level policies

macOS: Set the Settings key in the com.anthropic.claudecode managed preferences domain (via Jamf, Kandji, etc.) to:
{"disableAutoMode": "disable"}
Windows: Set the Settings key (REG_SZ) at HKLM\SOFTWARE\Policies\ClaudeCode (via Group Policy, Intune) to:
{"disableAutoMode": "disable"}