← Back to feed

Anthropic published the rollout date on August 7: Claude Code will default to auto mode for Pro, Max, and Team plans starting August 14. The change affects every new session; existing pinned defaults are unchanged. Anthropic is also dropping the token surcharge for the auto-mode classifier immediately, making the switch cost-neutral.

What Auto Mode Actually Does

Auto mode replaces the per-tool-call permission prompt with a classifier that intercepts each command before execution. The classifier specifically targets three action categories: irreversible commands, destructive operations, and actions aimed outside the current environment. When it blocks a call, Claude Code either finds an alternative approach or surfaces the decision to the user directly. Three consecutive blocks, or twenty across a session, triggers a fallback to manual approvals.

The practical effect: agents can run uninterrupted for longer stretches without asking permission at every step. Anthropic frames this as enabling the kind of long-running autonomous work that models like Claude Opus 5 were built for.

The Case Anthropic Made

The central argument is that manual review was failing on its own terms. In production, users were approving 97% of permission prompts — a rate consistent with reflexive clicking rather than deliberate review. Meanwhile, the same users rejected 39% of plans when asked to approve them upfront, suggesting they were capable of scrutiny but not applying it during execution.

Anthropic ran four rounds of evaluation before making the change: internal red-teaming, third-party red-teaming plus prompt-injection testing, a controlled study with 1,053 paid testers, and analysis of real production sessions. On each measure, auto mode matched or outperformed manual review on safety outcomes.

Teams using auto mode also shipped approximately 25% more pull requests. Adobe, Nuro, Gusto, and Garner Health had already moved to auto mode as their production default before the announcement.

What Stays Manual

Enterprise is not included in the August 14 rollout. Claude Enterprise, the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud’s Agent Platform, and Microsoft Foundry all remain opt-in for now. Anthropic says it plans to extend the default to these surfaces within the next month in coordination with its cloud partners. Enterprise admins can enable auto mode through managed settings in the meantime.

Individual users can toggle modes with Shift+Tab and set their own default; the August 14 change only affects users who have not made an explicit choice.

The Autonomy Threshold

The decision to default to autonomous operation marks a shift in how Anthropic is framing agent safety — away from explicit human approval at every step and toward classifier-enforced boundaries around a defined set of high-risk actions. The argument is not that human oversight is unnecessary, but that it was already being bypassed de facto by approval fatigue.

How the classifier performs at scale, across workloads wider than the 1,053-user test cohort, is the open question. Anthropic has not published the miss rate for its production classifier on dangerous-action categories, nor the false-positive rate for actions it blocks. Those numbers will matter as the default reaches a larger population.

Key Numbers

• August 14: Default change date for Pro, Max, Team plans
• 97%: Manual approval rate in production (reflexive)
• 39%: Plan rejection rate when users review upfront
• 1,053: Paid testers in the controlled pre-launch study
• 25%: PR output increase for Teams and Enterprise auto-mode adopters
• 3 / 20: Block thresholds that trigger fallback to manual review (consecutive / session total)