Anthropic shipped a meaningful update to Claude Code this week: auto mode, a new permissions setting that lets the AI make its own decisions about which file writes and bash commands to execute — without asking the developer for approval on each one.
The feature launched March 24 as a research preview on Team plans, with Enterprise and API access rolling out in the coming days. It works exclusively with Claude Sonnet 4.6 and Opus 4.6.
Claude Code's default behavior requires human approval before every file write and shell command — a deliberately conservative posture that becomes a serious friction point during long, complex tasks. The existing workaround, a flag called --dangerously-skip-permissions, removes all guardrails entirely. That's fine in a fully sandboxed environment; it's a liability everywhere else. Auto mode is the middle option that didn't exist before.
Here's how it actually works: before each tool call executes, a separate classifier model — running on Claude Sonnet 4.6 — reviews the action against the context of the conversation. Safe actions proceed automatically. Anything flagged as risky — mass file deletions, data exfiltration attempts, actions against unrecognized infrastructure, or behavior that looks like it's being driven by hostile content in a file or web page — gets blocked. If repeated attempts continue triggering blocks, the system escalates and prompts the developer for input. Claude's trusted zone is its local working directory and configured git remotes; everything else, including company source control systems and cloud storage buckets, is treated as external by default until explicitly whitelisted.
To enable it: claude --enable-auto-mode in the CLI, then toggle with Shift+Tab. VS Code and desktop users access it through Settings. Enterprise admins can disable it organization-wide via managed settings.
A few caveats worth noting. Anthropic recommends running auto mode in isolated environments — containers, VMs, or sandboxes — rather than directly against production systems. The classifier is a research preview, which means edge cases exist. Token usage, cost, and latency will increase modestly on tool calls due to the classification step running in parallel.
The broader context: this is part of a sustained push from Anthropic to make Claude Code a full-stack autonomous development tool. In recent weeks, the company also shipped Claude Code Review (automated bug detection before deployment) and Dispatch for Cowork (mobile task delegation to AI agents). Auto mode addresses the fundamental workflow tension at the core of agentic coding — the tradeoff between human oversight and uninterrupted execution — and takes a genuine step toward resolving it.
For developers who've been babysitting terminal approvals or quietly using --dangerously-skip-permissions because the default felt too slow, auto mode is the update worth enabling today.
Comments