New Delhi: Anthropic is changing the default way its AI coding assistant, Claude Code, operates. From August 14, the coding agent will use auto mode by default for Pro, Max and Team users, allowing it to make decisions and carry out tasks without repeatedly asking users for permission.
The change marks a significant shift towards more autonomous AI-assisted software development. Anthropic says its internal testing found that auto mode can identify potentially dangerous commands more effectively than humans manually reviewing permission requests. The company claims the system detected 89% of dangerous commands in testing, compared with just 11% identified through manual approval.
The move comes as AI coding agents increasingly take on longer and more complicated software-development tasks. Instead of simply generating snippets of code, these systems can now plan tasks, modify files, run commands, test changes and work for extended periods with less human intervention.
Claude Code will no longer ask for every approval
Claude Code previously required users to approve many actions before the agent could continue. Anthropic introduced an auto mode in March, allowing the coding assistant to make decisions independently for most tasks.
Now, the company is making that mode the default for a wider group of users.
Anthropic said Claude Code will switch to auto mode by default for Pro, Max and Team subscribers from August 14. Users will still be able to change the mode, while administrators will have options to control how the feature works within their organisations.
The change means developers will not have to repeatedly monitor permission prompts while Claude works through a lengthy coding task.
Anthropic says auto mode is safer
The most notable part of the announcement is Anthropic’s claim that allowing Claude Code to operate more independently can actually reduce risk.
According to the company, auto mode uses a separate classifier to evaluate commands and determine whether they could be dangerous. The system assesses actions for characteristics such as being irreversible, destructive or directed outside the user’s environment.
If a command is considered dangerous, Claude Code blocks it rather than executing it automatically.
Anthropic argues that this approach can be more reliable than asking users to approve every individual command.
The company said its testing showed auto mode identifying 89% of dangerous commands, while humans manually reviewing permission prompts identified only 11%.
Human approval may become a routine habit
Anthropic’s argument is based partly on how users interact with permission prompts.
The company said Claude Code users approve 97% of permission requests. This suggests that repeatedly asking users for permission may not necessarily provide strong protection if users become accustomed to simply clicking or approving requests.
The problem is similar to security warnings in other software. When people encounter too many warnings, they can become less attentive to individual alerts.
Anthropic believes its automated classifier can provide a more consistent layer of protection while allowing developers to work without constant interruptions.
Safety detection reportedly improves during longer sessions
Anthropic also reported that the difference between automated and human detection became larger during longer Claude Code sessions.
According to the company, human detection of dangerous commands fell to around 5% after 50 or more previous prompts, while auto mode’s detection rate remained stable.
The finding is particularly relevant as AI coding agents move towards longer-running workflows.
A developer working with an AI agent for several hours could encounter dozens or even hundreds of decisions. Requiring manual approval for every action could make such workflows slow and difficult to manage.
An automated safety layer allows the agent to continue working while still blocking commands that its classifier considers risky.
What happens when Claude detects a dangerous command?
Auto mode does not mean Claude Code receives unlimited freedom.
When the system detects a potentially dangerous command, it blocks the action. Claude can then attempt a safer alternative or ask the user for assistance.
Anthropic has also introduced additional safeguards for repeated blocked actions.
If Claude Code encounters three consecutive blocked commands, or reaches 20 blocked commands during a session, the system switches back to manual approvals.
This provides a fallback mechanism when the agent repeatedly encounters actions that the automated system considers unsafe.
Developers can run longer tasks
One of the biggest advantages Anthropic is highlighting is reduced interruption.
In traditional coding workflows, developers may need to remain nearby while an AI agent works because the system can stop and wait for approval.
Auto mode is designed to remove much of that “babysitting” requirement.
Anthropic says users can leave multi-hour tasks running in the background and return later to review the results.
This could be particularly useful for tasks such as fixing multiple bugs, running tests, updating documentation, refactoring code or preparing pull requests.
The broader direction is towards AI agents that can complete software-development workflows rather than simply assist with individual coding decisions.
Claude Code is being designed for autonomous work
The change comes as Anthropic increasingly positions Claude Code as an agent capable of handling long-running engineering projects.
Anthropic’s own 2026 research describes a shift from short coding tasks towards agents that can work for hours or even days while building and testing larger software systems.
The company has cited examples of Claude Code completing complex engineering tasks with minimal human intervention.
This means the software-development model is changing.
Instead of a developer asking an AI to write one function and then checking the result, an agent can increasingly be given a broader objective and allowed to plan and execute a sequence of tasks.
Auto mode works with long-running AI models
Anthropic said auto mode is particularly useful with models designed for extended tasks, including Opus 5.
Users can start one or several tasks in parallel and later return to completed pull requests for review.
This approach could allow engineering teams to delegate routine or time-consuming development work to AI agents while developers concentrate on higher-level decisions.
For example, an engineering team could assign an agent to work through a backlog of routine maintenance tasks while developers focus on architecture, product development and complex technical problems.
Companies are already using auto mode
Anthropic said several companies are already using auto mode as a production default.
The companies named include Adobe, Nuro, Gusto and Garner Health. Anthropic also said teams and enterprise users adopting auto mode ship around 25% more pull requests.
The figure indicates the potential productivity benefits of allowing AI agents to operate with fewer interruptions.
However, the reported productivity increase comes from Anthropic and should be viewed in that context rather than as an independently established industry-wide result.
Different companies may experience different outcomes depending on their engineering workflows, codebases and level of AI adoption.
Anthropic still warns about high-stakes changes
Despite describing auto mode as safer for most users, Anthropic has not suggested that developers should stop reviewing important changes.
The company specifically recommends that users continue to review high-stakes modifications to production infrastructure themselves.
This distinction is important.
A coding agent may be capable of identifying many dangerous commands, but automated safety systems cannot guarantee that every risky action will be detected.
Production environments can contain sensitive data, financial systems, customer information and critical infrastructure.
Human oversight therefore remains important when the consequences of an incorrect change could be significant.
Enterprise customers get more control
Auto mode will not immediately become the default for every Claude Code user.
From August 14, the default change applies to Pro, Max and Team users.
Enterprise customers will continue to have auto mode as an opt-in feature for now. The same applies to Claude API and several cloud platforms, including Amazon Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry.
Anthropic said it plans to make auto mode the default for these platforms in the following month after notifying administrators.
This phased approach gives organisations more time to evaluate the technology and establish internal controls.
Administrators can disable auto mode
Businesses will not be forced to allow autonomous operation.
Administrators can pin a preferred default mode or disable auto mode entirely. Individual users can also switch modes using Shift+Tab.
This could be important for companies operating in regulated sectors or environments where software changes require formal approval.
Organisations can therefore decide how much autonomy they are comfortable giving AI coding agents.
Auto mode classifier will not count towards usage limits
Anthropic also announced a change related to usage.
The additional tokens consumed by the auto mode classifier will no longer count towards users’ usage limits.
This means users should not have to sacrifice part of their normal Claude Code usage allowance simply because the safety classifier is operating in the background.
The change could encourage more developers to use auto mode for longer workflows.
The broader shift towards AI agents
Anthropic’s decision reflects a wider trend across the AI industry.
AI systems are increasingly moving beyond generating text or code in response to individual prompts.
Agentic systems can now plan multi-step tasks, interact with tools, execute commands and respond to changing conditions.
Software development is one of the areas where this shift is particularly visible because coding agents can directly interact with development environments.
The more autonomous these systems become, the more important safety mechanisms and permission controls become.
Why the change matters for developers
For developers, the biggest benefit could be time.
Constantly responding to permission prompts interrupts concentration and slows down long-running workflows.
An agent that can safely continue working without requiring approval for every routine action could make AI-assisted development more efficient.
Developers may instead spend their time reviewing completed work rather than monitoring every individual action.
This could fundamentally change how programmers interact with AI coding tools.
The risk of giving AI more autonomy
The shift also raises questions about how much independence AI agents should have.
An autonomous coding system has the ability to modify files and execute commands within its environment.
If the system makes a mistake, the consequences could be greater than those of a conventional chatbot generating incorrect code.
Anthropic’s automated classifier is intended to address this risk, but the company itself acknowledges that auto mode does not eliminate risk entirely.
For developers, the key issue will be balancing convenience with oversight.
Human developers are still important
The move towards autonomous coding does not mean developers are becoming irrelevant.
Instead, their role may shift.
Rather than manually writing every line of code, developers could increasingly focus on defining objectives, reviewing AI-generated changes, making architectural decisions and handling complex problems.
This could make software development more similar to managing a team of highly capable assistants.
The developer would remain responsible for deciding what should be built and ensuring that the final product meets technical and business requirements.
Security will remain a major concern
As AI agents gain more autonomy, security researchers and developers will need to pay close attention to how these systems interact with software environments.
A coding agent may have access to files, development tools, terminals and other resources.
That access can make the system highly useful but also increases the potential consequences of mistakes or malicious instructions.
Anthropic has previously described sandboxing and monitoring as important safeguards for Claude Code deployments. Its safety reports also discuss measures designed to reduce the risk of harmful actions by coding agents.
The challenge will be ensuring that these safeguards continue to work as agents become more capable.
What users should know before August 14
Pro, Max and Team users should expect Claude Code’s default behaviour to change from August 14.
Users will receive an in-app notice when the new default is introduced.
Those who prefer the existing permission-based approach will still be able to switch modes, while administrators will have additional controls for organisational deployments.
Developers working on critical production systems should also take Anthropic’s recommendation seriously and continue reviewing high-stakes changes manually.
What this means for the future of coding
Anthropic’s move is another sign that AI coding tools are moving towards greater autonomy.
The traditional model of asking an AI assistant for suggestions is gradually being replaced by systems capable of completing entire development workflows.
If automated safety systems can reliably identify risky actions, developers may become more comfortable delegating larger tasks to AI agents.
However, the success of this model will ultimately depend on how accurately these systems understand context and how effectively they handle unexpected situations.
Conclusion
Anthropic is making auto mode the default for Claude Code Pro, Max and Team users from August 14, marking a significant step towards more autonomous AI-assisted software development. The company says its testing found that auto mode detected 89% of dangerous commands, compared with 11% detected through human manual approval.
The company argues that automated review can be safer because users approve 97% of permission prompts, potentially making manual checks less effective over time. Auto mode instead uses a separate classifier to assess commands and blocks actions it considers irreversible, destructive or outside the user’s environment.
At the same time, Anthropic acknowledges that the technology does not remove all risks. Users are still advised to manually review high-stakes changes to production infrastructure, while enterprise customers will initially retain more control over whether auto mode is enabled.
The development signals a broader transformation in software engineering, where AI agents are increasingly expected to work independently for hours or longer. For developers, the future may involve less time approving individual commands and more time supervising, reviewing and directing increasingly capable AI coding agents.