The AI keeps asking for permission.
“Should I run this command?”
“Can I modify this file?”
“Do you want me to continue?”
“Can I execute this operation?”
Anthropic is now making a major change to Anthropic’s Claude Code that could make these interruptions far less common.
Starting August 14, Automatic Mode will become the default for new Claude Code sessions on Pro, Max, and Team plans.
Instead of asking the user for approval before every routine action, Claude will increasingly decide for itself whether an operation appears safe.
And the reason behind the change is particularly interesting:
Users approve almost everything anyway.
Claude Code Users Approve 97% of Permission Requests
According to Anthropic, users approve roughly 97% of Claude Code's permission requests.
That creates an interesting problem.
If an AI assistant constantly asks:
“Can I do this?”
the user will eventually stop thinking carefully about the question.
Some users already use options such as “always allow” or create permanent permissions for certain commands simply to avoid being interrupted.
Anthropic calls this phenomenon “approval fatigue.”
And it isn't unique to AI.
We have seen the same problem with antivirus warnings, browser permissions, smartphone notifications and other security prompts.
When users see too many warnings, it becomes increasingly difficult to distinguish a genuinely dangerous action from a routine one.
Eventually, the security mechanism itself becomes something people simply click through.
Claude Will Now Decide When It Needs to Ask
The new system doesn't mean Claude gets unlimited access to your computer.
Instead, Anthropic is introducing a classifier that evaluates tool calls and determines whether an action appears safe enough to perform automatically.
For example, if Claude needs to read a file, execute a routine command or perform another low-risk operation, it may simply proceed without interrupting the user.
But potentially dangerous actions will receive more scrutiny.
If an operation is destructive, irreversible or goes beyond the user's working environment, Claude can still stop and ask for explicit permission.
The philosophy is essentially:
Don't interrupt the user for routine actions. Interrupt them when something genuinely risky happens.
That sounds simple, but the numbers behind Anthropic's experiments are surprisingly dramatic.
AI Was Much Better at Spotting Dangerous Commands
Anthropic tested the system with 1,053 paid users.
During the experiment, researchers inserted an explicitly dangerous command among normal permission requests.
The automatic system blocked approximately 89% of the dangerous commands.
Human users, meanwhile, rejected only 13.6% of them.
That's an enormous difference.
But the most interesting part may be what happened as users became tired.
After users encountered 50 or more permission requests, their attention dropped significantly.
Their rejection rate fell to around 5%.
In other words:
The more often humans are forced to make security decisions, the less meaningful those decisions can become.
This is one of the strongest arguments for AI systems making routine security decisions themselves.
But Claude Isn't Perfect
There is an important catch.
An 89% blocking rate also means that roughly 11% of dangerous commands initially got through.
That's obviously not good enough for a system that could potentially control a development environment.
Anthropic says further testing with Apollo Research and an improved classifier reduced the miss rate to around 7%.
That's a significant improvement, but it still isn't zero.
And attackers are unlikely to sit still.
A malicious package, for example, could attempt to manipulate an AI coding agent into downloading and executing dangerous files.
This creates another category of risk involving software supply-chain attacks and AI agents.
So it would be a mistake to interpret the announcement as:
“Claude can now handle computer security by itself.”
It can't.
The more accurate interpretation is:
Claude is becoming better at deciding when human intervention is actually necessary.
What Happens When Claude Detects a Risk?
When Claude identifies an operation as potentially dangerous, it can block the action and, where possible, attempt a safer alternative.
If necessary, it can also ask the user for explicit approval.
There's another safety mechanism as well.
If the system repeatedly encounters rejected actions, it can return to the traditional manual approval process.
Enterprise teams can also establish stricter policies.
For example, organizations can define explicit permission rules for certain Git operations or create restrictions that users cannot simply override.
That means businesses can maintain much tighter control over what an AI coding agent is allowed to do.
The Biggest Benefit? Less Friction
For developers, the biggest advantage is obvious:
Claude can work for longer without constantly interrupting them.
That's important because Claude Code is designed to behave more like an AI developer working alongside you than a traditional chatbot.
Imagine telling an AI:
“Go through this project, fix the bugs, update the code and run the tests.”
If the AI asks for permission 50 times along the way, you're still doing a large part of the work yourself.
The whole point of an AI coding agent is to let it handle the routine steps.
Automatic Mode moves Claude Code closer to that vision.
Anthropic also says the automatic safety checks will not require an additional fee for Pro, Max and Team users.
This Could Be Bigger Than Claude Code
The most important part of this change may have little to do with Claude Code itself.
AI is gradually moving away from the “chatbot that answers questions” model.
The next generation of AI systems is increasingly becoming:
software that can actually do things on your behalf.
AI agents can browse websites, edit files, run commands, write code, test applications and interact with computers.
But there's a fundamental problem.
An agent that needs human approval for every single action isn't really autonomous.
Imagine asking an AI:
“Fix this project, find the bugs and make sure all the tests pass.”
If it asks you to approve 50 individual actions, the AI is effectively giving the work back to you.
That's why the industry is moving toward a different model:
More autonomy → better risk analysis → human intervention only when necessary.
The Bigger Question Is No Longer “Which AI Is Smarter?”
This is where things get really interesting.
As AI agents receive more control over computers, the most important question may no longer be:
“Which AI is the smartest?”
Instead, users may increasingly ask:
“Which AI can I safely trust with my computer?”
That's a much harder question.
Giving an AI more authority makes it more useful — but it also increases the potential damage if it makes the wrong decision.
Claude Code's Automatic Mode is therefore more than just a quality-of-life improvement for developers.
It is another step toward a future where AI doesn't simply tell us what to do.
It does the work for us.
And the real technological challenge will be figuring out exactly when we should let it act on its own — and when a human should take control.
For Claude Code, that future starts arriving on August 14.
Comments
Post a Comment