Anthropic just solved one of the most annoying problems in AI-assisted coding. As of March 24, Claude Code ships with auto mode - a feature that finally offers something between "approve every single file write manually" and "let the AI do whatever it wants."
The timing tracks with a major business milestone: Claude Code has crossed $2.5 billion in annualized revenue, more than doubling since January 1, 2026. Business subscriptions have quadrupled in the same period, and enterprise customers now account for over half of Claude Code's revenue.
How Auto Mode Actually Works
A separate AI model - not Claude itself - reviews every tool call before execution. Safe, routine operations like writing files and running tests get approved automatically. Destructive operations hit a wall.
The default block list covers:
- Mass file deletion (wiping directories in one shot)
- Sensitive data exfiltration (transmitting credentials, API keys, or personal data)
- Malicious code execution (privilege escalation, system damage attempts)
- Ambiguous high-risk actions where consequences are unclear
The classifier treats your local working directory and configured git remotes as trusted by default. Anything outside that - company source control systems, cloud storage, internal services - requires explicit admin approval.
When Claude tries a blocked action, it doesn't just stop. It attempts an alternative approach. If it keeps hitting the same wall, the system eventually surfaces a permission prompt for the human to decide.
Availability and Overhead
Auto mode is currently a research preview on Claude Team plans. Enterprise and API access is rolling out now. It works only with Claude Sonnet 4.6 and Opus 4.6 - older models and third-party platforms are excluded.
To enable it: run claude --enable-auto-mode in the CLI, or toggle it in VS Code settings. Once enabled, you can cycle between permission modes with Shift+Tab.
Anthropic is upfront about the tradeoffs. Every tool call now passes through an additional model evaluation, which adds "a small impact on token consumption, cost, and latency." They haven't published exact overhead numbers.
They're also clear about what this isn't: auto mode is not a sandbox. Anthropic still recommends running it in containers or dedicated dev environments, because the classifier can misjudge when context is ambiguous. Human review after long autonomous runs remains essential, especially for deletions or network calls.
For daily Claude Code users, this is a practical quality-of-life improvement. The old binary choice between constant interruptions and zero guardrails was the single biggest friction point. A smart middle ground was overdue.