Anthropic is making auto mode the default permission setting for Claude Code, allowing its AI coding agent to perform more actions without repeatedly asking users for approval as the company looks to balance greater autonomy with safeguards around potentially risky operations.
Starting August 14, auto mode will become the default for Claude Code users on Anthropic’s Pro, Max and Team plans. Users will still be able to switch to other permission settings if they prefer greater manual control over the agent’s actions.
Claude Code can perform tasks beyond generating or suggesting code. The agent can edit files, execute commands and interact with development tools, making permission controls an important part of how much autonomy users give the system.
Under conventional permission settings, Claude Code may ask users to approve actions before they are executed. Auto mode instead evaluates actions and decides when Claude can proceed independently and when additional scrutiny is required.
Anthropic introduced auto mode as a way to address what it describes as permission fatigue. According to the company, users approved roughly 93% of permission prompts in its telemetry, with attention to individual approvals declining as the number of prompts increased.
The company has argued that repeatedly placing a human in the approval loop does not necessarily provide effective oversight if users begin approving requests routinely.
Auto mode uses a model-based safety layer to assess actions before they are executed. Anthropic says the system catches roughly 83% of what it describes as “overeager behaviours” before execution, although the company acknowledges that probabilistic safeguards can still miss potentially harmful actions.
The change reflects a broader shift in AI coding tools from assistants that primarily generate code towards agents capable of carrying out multi-step development workflows with less direct supervision.
Claude Code’s increasing autonomy also raises questions about how AI developers manage the potential consequences of agents that can execute commands or interact with external systems.
Anthropic has been developing additional layers of protection alongside permission controls. These include sandboxing, virtual machines, network restrictions and controls over the external tools and information that an agent can access.
The company has said effective agent security requires overlapping safeguards rather than relying solely on either human approval or automated classification. Restricting what an AI agent is technically capable of accessing can limit the impact of an unintended action even when another safety mechanism fails.
The move also comes as Anthropic expands Claude Code’s role across professional software development. The company has added capabilities designed for longer-running coding tasks, code review and agentic workflows while making the tool available across multiple subscription tiers.
Competition in AI coding has intensified as OpenAI, Google, Microsoft and other technology companies develop tools capable of generating, reviewing and increasingly executing software development tasks.
For developers, Anthropic’s decision changes the default relationship between the user and the coding agent. Rather than requiring approval as frequently before acting, Claude Code will increasingly determine when it can proceed independently.
Users who require tighter controls can continue to select more restrictive permission modes. The effectiveness of auto mode will ultimately depend on whether it can reduce interruptions without introducing unacceptable risks as coding agents take on increasingly complex and autonomous tasks.