1 week ago
25
If you’ve been using Claude Code and feel like you spend half your time clicking “approve” on actions you were always going to approve anyway, Anthropic has noticed. The company is flipping auto mode to the default permission setting in its AI coding agent on August 14, replacing a manual approval system that, by the numbers, was mostly theatrical.
Under the old setup, users were approving roughly 93% of prompts. That’s not a safety gate. That’s a rubber stamp with extra steps.
What auto mode actually does
Auto mode sits between two extremes. On one end, you have the old approach: every action requires explicit human approval. On the other, full autonomy with no guardrails. Auto mode occupies the middle ground by using a classifier model to evaluate risk before executing any action.
The system works in two stages. First, a fast single-token filter screens incoming requests. If the filter flags something as potentially risky, a more thorough chain-of-thought mechanism kicks in to evaluate the action more carefully. Only user messages and tool commands pass through this classifier, which keeps the process lightweight.
Low-risk actions, like reading a file or running a standard build command, proceed without interruption. Higher-risk operations still trigger a manual approval prompt. Anthropic has also layered in allow/deny rules that let users customize what gets automated and what doesn’t. Organizations can set their own policies at the admin level.
From research preview to production default
Auto mode didn’t appear overnight. Anthropic first launched it as a research preview on March 24-25, 2026, giving developers several months to test the feature and surface edge cases before it became the standard experience.
The feature works across multiple deployment environments, including the Anthropic API directly, AWS Bedrock, and Google Vertex. Model compatibility extends to Sonnet 4.6 and Opus 4.6+, covering the range of Claude models developers are likely running in their coding workflows.
Administrators who prefer the old manual approval system can override the default.
The approval fatigue problem
The 93% approval rate under the old system tells a story about how humans interact with permission prompts. When nearly every request gets approved, the approval process stops being a meaningful safety check and starts being an annoyance. Auto mode attempts to solve this by only surfacing approval requests when the classifier identifies genuine risk, making those interruptions more meaningful when they do occur.
For developers spending hours in an IDE with Claude Code running alongside them, the practical difference is significant. Fewer interruptions means longer stretches of focused work. It also means the AI agent can chain together multi-step operations without stopping to ask permission at each stage, which is particularly useful for tasks like refactoring across multiple files or running test suites.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·