Starting next week, sessions across Claude’s web, desktop, and mobile apps will run in auto mode by default for users on Pro, Max, and Team plans, Anthropic announced on October 23 through the official @claudeai account. Rather than asking users to approve each step of an agentic task, an automated check will block risky actions instead — and anyone who prefers the old behavior can switch back to manual approval at any time. The change extends a shift that began in Claude Code, where auto mode became the default permission mode in August. According to Anthropic, its classifier-based checks caught 89 percent of dangerous commands in testing, compared with 14 percent under manual approval. The move removes the biggest friction point in letting Claude work on its own.
https://x.com/claudeai/status/2113360845409583307
What Changes For Pro, Max And Team Users
Auto mode changes how Claude handles multi-step tasks. Instead of stopping to ask for approval before each action — running a command, writing a file, calling a tool — Claude will proceed on its own and rely on an automated safety check to intercept the dangerous ones. The scope is the full Claude surface: web, desktop, and mobile apps all move to auto mode by default next week for Pro, Max, and Team subscribers.
The setting is opt-out, not opt-in. Users who want to keep approving every step can switch back to manual approval, so the announcement is a default change rather than a removal of the choice. Free-plan users are unaffected: the announcement names only the paid tiers. If you need a refresher on what the paid plans cover, see our breakdown of Claude’s Pro, Max, Team and Enterprise pricing and what each one unlocks.
The Automated Check Behind Auto Mode
The announcement pairs the default change with the mechanism that is supposed to make it safe. Auto mode reviews shell commands and actions with a separate classifier that flags dangerous operations — Anthropic’s quoted figures from testing show it caught 89 percent of dangerous commands, while manual approval caught only 14 percent. The comparison is pointed: the company is arguing that tired humans clicking “approve” on every step are worse at safety than an automated reviewer.
That logic explains the framing of the whole rollout. Anthropic is not claiming the classifier is perfect; it is claiming that per-step human approval was never a strong safety barrier to begin with. Users on the paid tiers get a faster, less interrupted agent, and Anthropic gets its strongest evidence yet for how it thinks agents should be governed — by machine review, not by nagging.
From Claude Code To Everywhere
The @claudeai announcement quotes the original Claude Code auto-mode post from August 7: starting August 14, auto mode became the default permission mode in Claude Code for Pro, Max, and Team users. The web, desktop, and mobile rollout is therefore the second stage of a deliberate sequence — the company tested the default change with developers first, measured the classifier’s catch rate against human approvers, and is now bringing the same behavior to every surface where people use Claude.
The timing is consistent with how Anthropic has been pushing Claude’s agentic capabilities forward in recent months: fewer approval dialogs, more autonomous runs, and safety carried by automated checks rather than by slowing the user down. Each step of that arc has widened the gap between what the assistant can do on its own and what it must ask permission for.
Why It Matters
Approval fatigue is one of the quietest problems in agent UX. A user who approves fifty routine steps stops reading them, and at that point the manual gate exists mostly as friction rather than protection. Anthropic’s numbers — 89 percent caught by the classifier against 14 percent by human reviewers — are its case that the friction was never buying much safety anyway.
If the default sticks, it changes the competitive posture too. Pro, Max, and Team subscribers will experience Claude as something closer to an autonomous coworker than an assistant that asks permission, and the rivals shipping their own agent modes now have a public benchmark for the safety claim behind it: don’t promise better approvals, promise better automated checks. The opt-out switch is the release valve — and its usage rate, when it emerges, will be the honest measure of whether users trust the classifier as much as Anthropic does.
Background
Anthropic has been building Claude’s agentic layer in stages: Claude Code brought the assistant into developers’ terminals, and permission modes governed what it could do unsupervised. Manual approval was the original default, a natural carryover from chat-era safety thinking. Auto mode inverts that: approve by default, block by exception. The August rollout inside Claude Code was the pilot; next week’s expansion to web, desktop, and mobile makes it the standard Claude behavior for paying users.
Source: @claudeai
0 Comments