Anthropic ships Auto Mode for Claude Code

Anthropic rolled out Auto Mode for Claude Code, a permission system that replaces per-action approval prompts with a model-based classifier that decides in real time whether a given command aligns with what the user asked for. It began as a research preview for Team-plan users in late March and had extended to Enterprise, the API and Max-plan users by mid-April, dropping its beta label along the way.

Anthropic’s own justification for building it was a measured statistic: before Auto Mode existed, users were already approving 93% of the permission prompts Claude Code showed them, indicating the prompts were mostly generating fatigue rather than catching genuinely risky actions. The system works in two stages — a prompt-injection probe screens tool outputs for attempts to hijack the model’s instructions, and a separate transcript classifier evaluates each proposed action against a set of criteria before letting it run. On real traffic, Anthropic reported the classifier’s false-positive rate (blocking safe actions) at 0.4% across 10,000 cases, and its false-negative rate on a smaller sample of 52 known overeager or risky actions at 17% — meaning roughly one in six dangerous actions in that sample still slipped through.

The launch reflected a broader shift in coding agents: as tools like Claude Code took on longer, more autonomous sessions, constant human-in-the-loop approval became a bottleneck rather than a safeguard, and labs began substituting automated judgment for manual review at the point of highest friction. By August 2026, Anthropic had made Auto Mode the default permission setting on its consumer and team plans.