Anthropic has launched auto mode in Claude Code, allowing the AI to autonomously handle multi-step software development tasks — including code generation, execution, and tool use — with reduced manual intervention. The feature replaces the previous permission-heavy model that caused approval fatigue. Auto mode uses a layered safety architecture: an input layer that inspects tool outputs for malicious content, an execution layer that evaluates each action before running it, and a two-stage classification pipeline that fast-tracks safe operations while escalating risky ones. Human approval checkpoints are preserved for sensitive operations, with a visual red spinner indicating when a permission check is triggered. The system also extends safety checks to subagent workflows, validating task alignment before delegation and reviewing execution history on completion to detect prompt injection. Anthropic acknowledges ongoing tradeoffs between safety and cost, and encourages users to remain aware of residual risks.
Sort: