Atlas
Plain EnglishWhat's new

Claude Code Starts in Auto Mode: What Teams Should Know

Claude Code now starts in auto mode by default. Here is what changed on 16 to 19 November 2026, what the safety check blocks, and how to switch it off.

What changed this week?16 November (2.1.326)17 November (2.1.327)19 November (2.1.328)

The short version

  • Claude Code now starts in auto mode in interactive sessions: it stops asking for routine approvals, and a separate AI model reviews actions before they run.
  • Claude Code 2.1.326 (16 November 2026) and 2.1.328 (19 November 2026) extended the default, including Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud, Foundry and the Claude apps gateway.
  • Auto mode blocks things like force pushes, production deploys and downloading and running code, but it is not a guarantee, so the finished work still needs a human review.
  • To keep the ask-first behaviour, set permissions.defaultMode to default in your personal settings file; an administrator can set permissions.disableAutoMode to disable for a whole company.
  • Fewer prompts means less of a built-in pause, so reviewing what the agent changed matters more.
On this page
  1. The short version
  2. What is auto mode in Claude Code?
  3. What changed this week?
  4. What does auto mode block, and what does it allow?
  5. What happens when the safety check blocks something?
  6. What does this mean for a team?
  7. How do I turn auto mode off or change the default?
  8. How Atlas handles this
  9. The bottom line
  10. Questions people ask

If you start Claude Code today, it will most likely begin in auto mode: it carries on with routine work without asking you to approve each step, and a separate AI model reviews actions before they run. That default has been widened twice in the past week. Claude Code 2.1.326 (16 November 2026) and 2.1.328 (19 November 2026) both touched it. This article explains what that means for a team, and what to do if you would rather keep the old behaviour.

What is auto mode in Claude Code?

Claude Code is an AI agent that reads and edits your software project. Its older habit was to stop and ask: "May I edit this file?", "May I run this command?" Those questions are called permission prompts. They are safe, but they are slow, and after the fortieth one most people just click yes.

Auto mode is Anthropic's answer. In its permission modes documentation, Anthropic says "a separate classifier model reviews actions before they run, blocking anything that escalates beyond your request, targets unrecognized infrastructure, or appears driven by hostile content Claude read." A classifier is simply a second model whose only job is to say allow or block.

This is a reduction in prompts, not a promise of safety. Treat it as a filter that catches a lot, not as a replacement for looking at what was done, especially for sensitive operations.

What changed this week?

Two things are worth separating: the default itself, and the tightening around it.

According to the same documentation, from Claude Code v2.1.283 onward auto mode is the built-in starting mode for interactive terminal and VS Code sessions on every plan and provider. The page does not say when 2.1.283 shipped, so treat that as background. What the Claude Code changelog dates inside the last week is this:

  • 16 November (2.1.326): terminal and VS Code sessions with no mode set start in auto mode, rather than Manual, when your organisation's policy has not loaded. Organisations using the Claude Code plus Cowork HIPAA configuration also start in auto mode by default, and the setting disableAutoMode turns that off. Auto mode also began ignoring some wildcard shell allow rules, so the safety check decides those commands instead.
  • 17 November (2.1.327): auto mode now ignores more allow rules that would run any package, such as one that lets the agent run an arbitrary package runner. The safety check decides those too. The same release changes a notice to say that an upcoming release removes a fallback to billed classifier requests; each action it would have checked will then be asked about or denied.
  • 19 November (2.1.328): auto mode was added for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway, and interactive sessions there with no permission mode set start in it. The same release makes plan mode and command prompts on those models offer auto mode, and disableAutoMode turns both off.

In short: the default now reaches more setups, and the rules around it got stricter at the same time.

What does auto mode block, and what does it allow?

The documentation gives these examples of what the classifier blocks by default:

  • downloading and running code, such as curl | bash
  • force pushes
  • production deploys and migrations
  • mass deletion on cloud storage
  • granting access permissions on cloud accounts or repositories

And it gives these examples of what it allows by default: local file changes in your working folder, installing dependencies declared in your project's lock files or manifests, read-only web requests, and pushing to a branch of the repository you are working in. Pushing to branches with deploy-like names, such as production or gh-pages, is an exception.

These are examples, not a complete list, so read the documentation page before relying on a specific case.

What happens when the safety check blocks something?

Claude Code shows a notification and lists the action under a "Recently blocked" tab in its permissions screen, where you can approve it so Claude may retry. If the check blocks an action three times in a row, or 20 times in total, auto mode pauses and Claude Code goes back to asking you. Approving the prompted action resumes auto mode. Anthropic says those thresholds cannot be changed.

If a session runs without anyone at the keyboard (a -p run, which is Claude Code used from a script) and there is no way to ask, then when repeated blocks reach the threshold the action does not run and Claude keeps working. The run does not stop. That matters for anyone leaving an agent overnight: a block can quietly change the result, so you need to read the final report. Note also that on version 2.1.302 and later, a -p run with no permission mode set starts in Manual mode, not auto, and calls that need approval are denied unless a rule or hook allows them.

What does this mean for a team?

Faster, quieter sessions. Fewer interruptions means people can hand the agent a task and come back to a finished result. That is the point.

Less of a natural checkpoint. A prompt used to force a human to glance at what was about to happen. Now that glance is gone for routine actions, so the review of the finished work has to carry more weight.

Check which model and platform you use. The documentation says older models, including Sonnet 4.5, Opus 4.5 and Haiku 4.5, are not supported on any provider. If auto mode is selected but unavailable, Claude Code starts the session in Manual mode instead. So two people on the same team can have different experiences. The documentation's list of supported models for Bedrock, Google Cloud and Foundry names newer models than the 2.1.328 changelog entry does, so check the current page for your provider.

Be careful with your own allow rules. Since 16 and 17 November, broad rules such as "always allow this package runner" are ignored in auto mode. If your team relied on one, expect new blocks or questions.

Remember what a classifier can miss. Hostile text hidden in a web page or a file is one reason the check exists, but it is not a wall.

How do I turn auto mode off or change the default?

Per the documentation, a person can set the starting mode for terminal sessions in their personal settings file, ~/.claude/settings.json, using the permissions.defaultMode key. Setting it to default starts every terminal session in Manual mode, the ask-first behaviour:

{"permissions": {"defaultMode": "default"}}

A single session can also be started with claude --permission-mode auto (or another mode), and Shift+Tab switches the mode inside a running session.

Two details catch people out. Putting "auto" in a project's .claude/settings.json does not take effect. And in VS Code the claudeCode.initialPermissionMode setting does not accept auto; to start in auto there, leave it unset.

For a whole company, an administrator can set permissions.disableAutoMode to "disable". That removes auto mode from the Shift+Tab cycle, and a session started with --permission-mode auto starts in Manual instead.

How Atlas handles this

Atlas is a team of AI engineers built on Claude Code, and each engineer is its own Claude Code session. Atlas keeps five decisions for you whatever else happens: production go-lives, spending, deleting data, sending anything outside, and dropping scope. It also has a "Run unattended" switch on the chat's top bar; while it is on, questions meant for you go to PA, your project assistant, and those five decisions still wait for you.

Atlas does not yet show which permission mode each engineer is in, or tell you when auto mode pauses and a session waits for a prompt. That is coming to Atlas, built in, so it needs no setup on your side. Until then, check the mode in your own Claude Code sessions as described above. Atlas is an independent product. You can read more on the Atlas home page.

The bottom line

Auto mode is now the normal way interactive Claude Code sessions start, and the last week made it both more widespread and stricter. It saves real time, but it is a convenience with a safety net, not a guarantee. Decide on purpose: keep it for low-risk work, switch the default off where the stakes are high, and spend the minutes you save on reviewing what was actually built.

Questions people ask

Does Claude Code start in auto mode now?

In interactive sessions, mostly yes. Anthropic's documentation says that from Claude Code v2.1.283, interactive terminal and VS Code sessions start in auto mode on every plan and provider. Versions 2.1.326 (16 November 2026) and 2.1.328 (19 November 2026) extended this to more cases. Scripted -p runs start in Manual on 2.1.302 and later.

Is auto mode safe?

It is safer than approving nothing, but not a guarantee. A separate model reviews actions and blocks things like force pushes and production deploys, but it only reduces prompts. You should still review sensitive operations and the finished work yourself.

How do I stop Claude Code from starting in auto mode?

Set permissions.defaultMode to default in your personal settings file, ~/.claude/settings.json. An administrator can set permissions.disableAutoMode to disable in managed settings to remove auto mode for everyone.

What happens when auto mode blocks an action?

Claude Code shows a notification and lists the action under Recently blocked, where you can approve it. After three blocks in a row or 20 in total, auto mode pauses and Claude Code goes back to asking you. In a script with no one to ask, the action just does not run.

Does auto mode work on Amazon Bedrock or Google Cloud?

Yes for some models. Claude Code 2.1.328, dated 19 November 2026, added auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway. Older models such as Sonnet 4.5 are not supported.

Sources

  1. Claude Code docs: Permission modes (auto mode)
  2. Claude Code changelog (2.1.325 to 2.1.328)

Researched and drafted with Claude, then checked against the sources above. Atlas is an independent product, not affiliated with Anthropic or OpenAI. Claude and Claude Code are trademarks of Anthropic; Codex is a product of OpenAI.