Claude Code in August 2026: auto mode by default, and what that means for your team

Auto mode became the default for new Claude Code sessions on Pro, Max, and Team plans on August 14, 2026. Before that it was opt-in.

If you manage a team on Claude Code, this changed the default security posture of everyone's terminal, and it is worth understanding precisely rather than approximately.

What auto mode is

Auto mode is a permission system. Rather than prompting you before every action Claude wants to take, an AI classifier evaluates each command and automatically approves the ones it judges safe, while still blocking ones it judges dangerous.

The intent is obvious to anyone who has used Claude Code for a long session: the prompt fatigue is real, and a developer who has approved four hundred consecutive ls commands is not meaningfully reviewing the four hundred and first.

What changed, exactly

Default now: Pro, Max, and Team plans. New sessions run in auto mode.

Still opt-in: Claude Enterprise, the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry.

That split is the interesting part. Anthropic changed the default for individual and small-team plans, and left it unchanged for every enterprise and cloud-marketplace surface. Read that as an explicit judgement: auto mode is right for a developer working on their own repository, and not something to impose on an organisation that has a security review process.

If you previously set a different default, you may get a one-time prompt asking whether to switch. Pinned defaults are not changed.

Turning it off

Per session: Shift+Tab in the CLI cycles modes, or use the mode dropdown in the desktop app.

Org-wide, as an admin: managed settings expose two controls.

• defaultMode — set the mode new sessions start in
• disableAutoMode — remove auto mode as an option entirely

The second is the one to reach for if you have a compliance answer that depends on a human approving each action. defaultMode is a default; a developer can still switch. disableAutoMode is a policy.

Anthropic's own caveat

From the announcement: it relies on classification systems and therefore does not eliminate risk. Anthropic recommends manual review for high-stakes production infrastructure changes.

That is an accurate framing and worth taking literally. A classifier is a model making a judgement, and models are wrong sometimes. The question is not whether auto mode is safe in general — it is whether it is safe on the specific machine, with the specific credentials, in the specific repository.

Where to turn it off, specifically

Rather than an org-wide on or off, the useful decision is per-context:

Leave auto mode on for local development repositories, personal projects, scratch work, anything where the worst case is a bad commit on a branch, and any long refactor where prompt fatigue would otherwise make you approve without reading.

Turn auto mode off on any machine with production credentials in the environment, in infrastructure-as-code repositories where an apply is one command away, wherever the CLI has access to a live database, and in shared or CI environments where the blast radius is not one person's afternoon.

The failure mode to avoid is a developer who has auto mode on in their normal repo, switches to the infra repo, and does not notice the mode carried over. If your organisation has an infra repo, disableAutoMode scoped to those machines is a cheaper control than remembering.

The rest of the month

Two other Claude Code changes landed in August 2026.

Self-hosted environments (August 6, public beta). Claude Code sessions can run in an environment you host rather than Anthropic's. For organisations that could not adopt cloud-run sessions for data-residency reasons, this is the unblock. It pairs with the self-hosted sandboxes for Managed Agents that shipped in June.

Workbench became Playground (August 18). In the Claude Console, Workbench is now Playground. It supports every Messages API parameter, ships templates demonstrating code execution and web search, and shows the full SDK request alongside the API response for each run. The old Workbench and prompt tools APIs retired on August 17, so if you had automation calling those, it has already broken.

What to do

1. Check which mode your team is actually in. The default changed underneath them; most people will not have noticed.
2. Decide your infra-repo policy and enforce it with disableAutoMode rather than a reminder in a Slack channel.
3. If you are on Enterprise, nothing changed — but decide deliberately whether to opt in rather than leaving it unexamined.
4. Re-read your Claude Code antipatterns. Several of them get more likely, not less, when the approval prompts stop.

Related: Claude Code for your team · June 2026 updates · May 2026 updates · Anthropic's effort level guide