In brief: Anthropic is making Auto Mode the default in Claude Code: a classifier automatically checks agent actions, blocks dangerous commands, and enables uninterrupted sessions nine times longer than before, according to internal tests.
Anthropic has enabled Auto Mode as the default in Claude Code: instead of approving every action individually, a classifier decides which commands are potentially dangerous and blocks them. With this, Anthropic aims to ease the classic trade-off between speed and safety in agentic coding.
Previously, every action proposed by an agent in Claude Code had to be manually confirmed. The new Auto Mode replaces this approval step with a classifier that evaluates each planned action and automatically blocks potentially harmful commands. According to Anthropic, in internal evaluations the classifier identified more dangerous actions than developers who manually clicked through approval prompts. The classifier’s robustness was additionally verified through external red-teaming. One measurable effect: according to Anthropic, Claude Code sessions now run on average nine times longer without interruption compared to the previous default setting.
For engineering teams, this fundamentally changes how agentic coding is handled in practice. Permission fatigue — constantly clicking through approval dialogs — had previously been a major bottleneck for longer-running or parallelized coding sessions. At the same time, fully disabling approval checks carries risks such as prompt injection, scope drift, or accidental deletion of production resources. Auto Mode is designed to close this gap without requiring teams to give up control entirely.
Anthropic cites the company Nuro, which works on Level 4 autonomous driving technology, as an example. Staff Software Engineer Kai Zhou had already built his own mechanism before the official Auto Mode was introduced: a hook that sent pending actions to a small model, automatically approved routine tasks, and forwarded sensitive cases to Slack for review. After the official Auto Mode rollout, he set this project aside. He now uses Auto Mode for all of his coding work and frequently runs three to four parallel sessions. The exception is work affecting other teams, such as Claude Code reviewing pull requests — in these cases, he switches back to interactive mode.
At Nuro as well, Auto Mode does not run without limits: particularly risky commands, such as recursive delete operations, are fundamentally blocked in the settings, and the classifier makes its decisions within these guardrails. According to Nuro, the additional benefit lies primarily in tasks with a clear, measurable success metric — for example, iteratively improving evaluation metrics for autonomous driving software overnight. This allows an agent to independently analyze misclassifications, develop proposals, run experiments, and deliver finished pull requests by the next morning. At Gusto, an SMB software provider, the switch to Auto Mode was likewise driven by security considerations, aiming to reduce permission fatigue without sacrificing control.
Source: claude.com · Published August 6, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.