On 24 March 2026, Anthropic launched auto mode for Claude Code as a research preview. The feature lets Claude decide which terminal commands, file edits, and tool calls are safe to execute without prompting the user for permission — a significant step towards fully autonomous agentic coding.
Auto mode works through a safety classification system. Claude evaluates each proposed action against a set of risk heuristics: read-only operations, file edits within the current project, and standard build commands are classified as safe and executed immediately. Destructive operations like force-pushing to git, deleting files outside the project boundary, or running unknown shell commands still require explicit user approval. The system errs heavily on the side of caution, with Anthropic stating that false positives (asking when it could have acted) are preferable to false negatives.
For context engineers and Claude Code power users, auto mode dramatically reduces the friction of long-running agentic sessions. Tasks that previously required dozens of permission confirmations — like refactoring across multiple files, running test suites, and committing results — can now flow without interruption. The feature pairs naturally with the recently shipped computer use and voice mode capabilities.
Anthopic emphasises this is a research preview with guardrails. Users can configure allowed and blocked action patterns, and the system logs every autonomous decision for review. The goal is to find the right balance between developer productivity and safety — letting the agent move fast on routine operations while keeping humans in the loop for consequential decisions.
