On August 10, 2026, Beijing time, TechCrunch AI's public information stream reported that Anthropic is enabling auto mode by default for Claude Code and summarized the change as requiring less human supervision for programming. The item was published at 19:20:32 UTC on August 9. Anthropic's current permission documentation does not equate default visibility with default startup, so the two states need to be distinguished: eligible accounts will see Auto in the mode selector and no longer receive a confirmation prompt when switching, while newly created sessions with unchanged settings still start in Manual.
Different Interface Entries for Auto Mode
The current documentation separately explains the entries for CLI, VS Code, JetBrains, Desktop, web, and mobile interfaces. In CLI, Shift+Tab can be used to cycle through modes; in VS Code and Desktop, 'Auto' is displayed when the account meets the conditions; cloud sessions can also select from a drop-down menu of modes. Officially listed providers include Anthropic API, Claude Platform on AWS, Amazon Bedrock, Google Cloud Agent Platform, and Microsoft Foundry, with each provider still needing to use models listed in the documentation. Remote Control sessions only display the current mode of local sessions within the application and cannot select Auto or Bypass permissions from this interface.
Default availability and default startup are two different product states. Whether a new session directly enters auto mode depends on the value set for permissions.defaultMode in user or hosting settings; only setting this to 'auto' will start the session in that mode. If the corresponding value remains as Manual, users will simply see an additional Auto entry among selectable modes without automatically changing their session state due to increased product availability. Official documentation also prohibits project-level settings from granting this default on their own to prevent code repositories from altering startup permissions through their configurations.
Open Timeline and Approval Fatigue
Anthropic first released auto mode as a research preview for the Team plan on March 24, 2026. The same official announcement was updated on July 10 to state that this mode is now available to all users. Current documentation notes that from Claude Code versions 2.1.158 through 2.1.206, some cloud providers require an old enable environment variable; starting with version 2.1.207, the variable remains only for compatibility and no longer has any enabling effect. The official page does not list a specific version or precise release date for changes mentioned on August 9.
Anthropic positions auto mode as a solution between manual approval of each action and skipping all permission checks entirely. In the former case, actions such as file writes or command executions wait for human confirmation; in the latter, --dangerously-skip-permissions bypasses these checks, which Anthropic advises should not be used outside isolated environments. The company's engineering article states that Claude Code users approve 93% of permission prompts, with manual approval leading to approval fatigue and long-running tasks being interrupted by waiting for confirmation. This 93% statistic is based on data from Anthropic’s own products; public articles do not disclose the sample size, time period, account types, or external review results.
Dual-layer Inspection and Explicit Permission Rules
The official engineering article divides the workflow into two layers. On the input side, server prompt injection probes scan file contents, web pages, command outputs, and responses from external tools; suspicious content is flagged with warnings before entering the context. The session classifier on the output side checks user requests and tool calls before actions are executed; Anthropic says it can prevent actions that exceed the request scope, point to unknown infrastructure, or are driven by suspicious external content. According to the current documentation, work directories and repository remotes configured when sessions start are trusted by default; other targets require classification, and remotes added or redirected after startup do not automatically enter the original trust boundary.
auto mode does not eliminate explicit permission rules. When an operation matches a permissions.ask content rule, the system still forces an approval prompt that the classifier cannot bypass; permissions.deny blocks actions before the classifier runs. Team and Enterprise administrators can disable auto mode in managed settings by setting permissions.disableAutoMode to disabled; Auto is then removed from the mode selector, and attempts to specify this mode through startup parameters are rejected.
"Default On" Source Boundaries
There are differences in the expression of 'default' that need to be preserved between two types of sources. The TechCrunch headline uses 'default on,' while Anthropic's current documentation separately describes default visibility, no longer requiring confirmation when switching, and the launch mode for new sessions. Only the last item directly determines whether a session starts without manual approval each time. The March announcement still retains the old steps that require manual activation, but it has added an update about full availability in July on the same page; historical activation instructions cannot replace current permission documentation.
As of August 10, 2026, Beijing time, Anthropic has not publicly disclosed the independent version number corresponding to the changes mentioned on August 9, the proportion of phased accounts, platform coverage progress, or the time of enabling. Existing sources have also not provided completion rates for tasks after the change, false rejection rates, or release rates for risky operations. The official documentation clearly states that auto mode will only reduce permission prompts and does not guarantee safety; classifiers may still allow risky operations or block harmless ones. Anthropic recommends using this mode only when users trust the general direction of the task and not as a substitute for review in sensitive operations; publicly available materials also contain no independent product testing.