Anthropic is set to make auto mode the default permission setting in Claude Code beginning August 14, replacing a manual approval system that, according to the company's data, users rarely declined.

Under the previous configuration, users approved approximately 93% of all prompts. Anthropic concluded that this rate rendered the manual approval step largely ineffective as a safety mechanism. The finding aligns with a well-documented phenomenon in security and interface design known as alert or approval fatigue, where habitual approval patterns erode the value of permission prompts regardless of the specific action being requested.

How Auto Mode Works

Auto mode is designed as a middle ground between two extremes: requiring explicit human approval for every action on one end, and granting the AI agent full autonomy with no guardrails on the other.

The system evaluates risk before executing any action using a classifier model that operates in two stages. A fast single-token filter first screens incoming requests. If that initial filter flags a request as potentially risky, a more thorough chain-of-thought mechanism is activated to assess the action in greater detail. Only user messages and tool commands pass through the classifier, which keeps processing overhead minimal.

Low-risk actions — such as reading a file or running a standard build command — proceed without interruption. Higher-risk operations still trigger a manual approval prompt. Anthropic has also introduced allow/deny rules that let users customize which actions are automated and which require approval. Organizations can enforce their own policies at the admin level.

From Research Preview to Production Default

Anthropic first introduced auto mode as a research preview on March 24–25, 2026, giving developers several months to test the feature and identify edge cases before it became the standard experience.

The feature is compatible across multiple deployment environments, including the Anthropic API directly, AWS Bedrock, and Google Vertex. Model support covers Sonnet 4.6 and Opus 4.6+, encompassing the Claude models developers are most likely to use in coding workflows.

Administrators who prefer the previous manual approval system retain the ability to override the new default.

Addressing Approval Fatigue

The 93% approval rate under the old system highlights a broader issue with how users interact with permission prompts. When nearly every request is approved, the approval process ceases to function as a meaningful safety check and becomes a procedural obstacle.

Auto mode aims to resolve this by surfacing approval requests only when the classifier identifies genuine risk, making those interruptions more deliberate when they do occur.

The shift comes as AI coding tools across the industry — including GitHub Copilot, Cursor, and others — continue to expand the scope of actions their agents can take within developer environments, making permission architecture a key differentiator in how much trust teams place in autonomous coding workflows.

For developers working in an IDE with Claude Code active over extended sessions, the change is notable. Fewer interruptions allow for longer periods of focused work. The AI agent can also chain together multi-step operations without pausing for permission at each stage, which is particularly useful for tasks such as refactoring across multiple files or running test suites.