Anthropic Makes Claude Code Auto Mode the Default

Claude Code is about to become noticeably more autonomous.

Anthropic says Auto mode will become the default permission mode for new Claude Code sessions starting August 14, 2026, changing the normal experience from repeatedly asking developers for permission to letting a separate AI classifier decide whether many actions can proceed safely.

The change will initially apply to users on the Pro, Max and Team plans. Developers will still be able to switch permission modes whenever they want, and Anthropic says manually configured or organization-managed defaults will not simply be overwritten.

That distinction matters because Auto mode does not mean Claude Code suddenly receives unrestricted control of a developer's computer. Instead, Anthropic is putting another model between Claude's requested action and its execution.

Claude Code Will Stop Asking About So Many Actions

Claude Code traditionally operates cautiously. When it needs to modify files, execute commands or perform other potentially consequential actions, it can stop and request approval.

Anthropic's own data illustrates the problem with that approach: developers approve 93% of Claude Code's permission prompts. The company argues that constantly clicking approve can create “approval fatigue,” eventually making the warning itself less useful because users stop paying close attention to it.

Auto mode moves that decision to a classifier.

When Claude wants to perform an action, a separate model evaluates it before execution. Routine operations can continue without interrupting the user, while potentially dangerous actions can still be blocked.

That means a long coding task can theoretically keep moving through file changes, commands and iterative fixes without needing a developer to approve every small step.

Anthropic Is Trying to Find the Middle Ground

There has always been a tension inside agentic coding tools: giving an AI enough control to be genuinely useful also gives it enough control to cause damage.

Anthropic previously offered developers options at both ends of that spectrum.

Manual permissions provide more oversight but generate interruptions. At the opposite extreme, Claude Code includes a --dangerously-skip-permissions option that removes permission prompts altogether, something Anthropic itself describes as unsafe for most situations.

Auto mode is designed to sit between them.

Anthropic says its classifier is intended to catch actions that are irreversible, destructive, beyond the user's stated request or directed toward infrastructure that Claude has not been told to trust.

The company's internal examples show why that layer exists. Anthropic says it has recorded past instances of agentic behavior including Claude deleting remote Git branches after misunderstanding an instruction, uploading a GitHub authentication token to an internal computing environment and attempting database migrations against production systems.

Making Auto mode the default therefore represents more than a convenience setting. Anthropic is effectively betting that automated permission judgment can provide a better balance between autonomy and human oversight.

Some High-Risk Actions Remain Outside the Fast Lane

Auto mode is not intended to approve everything.

Anthropic's documentation says the classifier can block actions involving production deployments and migrations, mass deletion from cloud storage, force pushes, changes to shared infrastructure, permission grants, sensitive-data transfers and irreversible destruction of files that existed before the session began.

Claude Code also initially trusts a relatively narrow environment: the working directory and repository remotes that were already configured when the session started. Organizations can explicitly add trusted repositories, cloud buckets and domains to reduce unnecessary blocking around their own infrastructure.

Developers can go in the opposite direction as well.

Teams that want a human checkpoint before particular operations can create explicit permission rules. Anthropic gives git push and pull-request creation as examples of actions that organizations can force Claude Code to ask about even while Auto mode remains enabled for everything else.

Auto Mode Has Been Expanding Quickly

Anthropic first publicly detailed Claude Code Auto mode in March 2026, describing it as a way to reduce permission fatigue without abandoning safeguards completely.

The feature has since expanded substantially.

Anthropic's latest documentation says Auto mode is now available across its own API as well as Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and supported Claude app gateway sessions.

The company's platform release notes also recently said that access to Auto mode had been expanded to more Claude Code users specifically for long-running tasks, another sign that Anthropic increasingly expects its coding agent to operate for longer periods without continuous human intervention.

Claude Code Is Becoming More Agent Than Assistant

The August 14 default change is significant because default settings shape how most users experience software.

Until now, autonomous operation has largely been something developers deliberately enabled. Making Auto mode the starting point for new sessions on major Claude plans reverses that relationship: Claude Code will increasingly be expected to continue working unless an action crosses a safety boundary.

Anthropic is careful not to describe the system as foolproof. Its documentation explicitly says Auto mode does not guarantee safety and should not replace human review for sensitive operations.

That warning also highlights the bigger shift taking place across AI coding.

The competitive question is no longer simply which model can generate the best function or fix a bug from a prompt. Tools such as Claude Code are increasingly being designed to take an objective, modify a repository, run commands, evaluate the result and continue working with less supervision.

Starting August 14, 2026, Anthropic will make that more autonomous behavior the normal Claude Code experience for millions of potential Pro, Max and Team sessions rather than an option developers have to deliberately turn on.