The report
Anthropic's product guidance describes auto mode as reducing repeated manual permission prompts while screening risky or irreversible tool actions. An October 22 announcement said that starting the following week it would become the default for certain sessions on Pro, Max and Team plans, across supported web, desktop and mobile clients. Manual approvals remain an available option; rollout, feature scope and administrator controls must be checked for a particular deployment.
The design changes where attention is spent. A tool-permission preflight can stop a class of disallowed calls, but an accepted call may still contain the wrong beneficiary, file path, amount or recipient if those semantics are not enforced at the point of effect. This announcement is a product-policy change, not a reported autonomous-agent breach. It also differs from Claude Duet's preview commitments to obtain separate approval for consequential sends, payments and deletions. Anthropic user guidance.
Approval becomes a policy decision rather than a count of prompts
Auto mode shifts many routine action checks from repeated human confirmations to a tool-action risk decision. The policy is meaningful only if it checks the actual operation against the current task and retains a separate boundary for consequential effects, including money movement, publishing and deletion. A lower prompt count is not a measured reduction in unauthorized executions. The October 22 statement concerns a planned rollout, not proof that every named client and plan had already switched. Announcement; permission guidance.
What "automatic" authorizes
The product announcement concerns a default approval mode, not a blanket permission for every tool effect. It claims a classifier can intercept risky actions, but it does not publish a denominator of dangerous actions attempted, actions blocked incorrectly, or harmful actions that passed the check. A user-facing override to manual approvals provides an alternative workflow, not retrospective evidence that previous automatic actions were safe. Assessment should separate reversible document navigation from payment, publication and privileged system modifications; the announcement does not demonstrate a reliable failure rate for any of those categories. Product guidance; dated announcement.