Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 7 min read

Claude Code 2.1.328 Brings Auto Mode to Bedrock and Foundry

Claude Code 2.1.328 (November 19) adds auto mode on Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry Interactive sessions on those three with no permission mode set now start in auto mode, not manual p

  • Claude Code 2.1.328 (November 19) adds auto mode on Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry

  • Interactive sessions on those three with no permission mode set now start in auto mode, not manual prompts

  • Plan mode and Bash prompts on third-party providers now offer auto mode too, one setting turns both off

  • /code-review on most models now flags only bugs and CLAUDE.md violations, cleanup moved to /simplify

On November 19, 2026, Anthropic shipped Claude Code 2.1.328, and the change worth stopping for is auto mode leaving its home turf. Auto mode, the permission mode that took over as the default for Pro, Max and Team plans back in August, now runs the same way when Claude Code is routed through Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry. Until this release, a session on any of those three kept prompting for almost everything, the behavior this blog covered when the default flipped on the consumer plans.

Three providers, one version number. That's the whole shape of the change.

What changed in Claude Code 2.1.328?

The release note is specific: auto mode is added for Claude Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway, and interactive sessions with no permission mode set now start in it. That second clause is the one to read twice. It is not an opt-in you discover in settings. If you have never touched your permission mode on one of those providers, your next session opens already running the classifier that decides what gets through without a prompt.

There is a second, related change tucked into the same release. Plan mode and Bash prompts for Opus 4.6 and Sonnet 4.6 on third-party providers now use and offer auto mode the same way newer models already do elsewhere. One setting, disableAutoMode, turns both of those off if you want the old prompting back. That's the whole opt-out. No migration wizard, no per-action toggle, just a single flag.

I read this against the August rollout rather than taking the changelog line alone, because the two releases are clearly the same project shipped in stages. Back then, auto mode covered the CLI, IDE extensions, the desktop app and claude.ai/code, and it required Opus 4.6, Sonnet 4.6 or Fable 5. The model requirement hasn't moved. What moved is the list of places that requirement actually gets you auto mode. Bedrock, Google Cloud's Agent Platform and Microsoft Foundry were the gap, and 2.1.328 closes it.

A second, smaller fix in the same release matters if you already tried auto mode on one of these providers and it misbehaved. A bug made auto mode fail every turn outside Claude Desktop, when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS was set behind a gateway that rejects beta flags. That's fixed in 2.1.328. If that was you in the last few weeks, it's worth trying again now rather than assuming it's still broken.

Does auto mode behave differently on a third-party provider?

Not in the part that decides what runs. The classifier is the same one I wrote about in August: it looks at a shell command, a file write outside your working directory, or a network call, and either lets it through or blocks it, instead of a prompt firing for almost everything the way manual mode does. The three-strikes fallback is also unchanged. Three blocked actions in a row, or twenty in one session, drops you back to manual prompting until you approve something again.

What differs is who decided your default. On Pro, Max and Team, Anthropic flips the switch and gives you a one-time prompt you can decline. That prompt doesn't exist on the other side.

On Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, the account usually belongs to whoever set up the cloud project, not to you personally, and disableAutoMode is the lever that account controls. If you're a solo builder working inside someone else's AWS or Azure tenant, a client's or an employer's, you don't get the same one-time opt-out prompt. The session just starts in auto mode the next time you run it, unless that setting was already flipped.

Nobody asks you first.

Token billing is the other place this matters. The August piece on this blog noted that the classifier's extra round trip isn't billed on Pro, Max or Team, only on Enterprise and on API, Bedrock-style routes. Nothing in 2.1.328 changes that line. If you're paying per token on one of these three providers, the safety layer is now running there by default, and so is its token cost. That's a real number on an invoice, not an abstraction.

What else shipped in 2.1.328 that's worth knowing?

Two smaller changes affect daily use regardless of which platform you're on. /code-review on most models now reports only bugs and violations of your project's CLAUDE.md file. Reuse, simplification and efficiency suggestions, the stuff /code-review used to mix in, moved to a separate /simplify command. If you ran /code-review as a catch-all before pushing, you now need both commands to get what one used to give you.

Each finding from /code-review is also marked Important, Pre-existing or Nit now, and a clean review says so in plain words instead of an empty list you have to interpret yourself. Small, but it saves a second of guessing every time.

/bug asks a plain yes or no before it sends a report now, so holding Enter through a stack of prompts doesn't fire one off by accident anymore.

I've done exactly that, more than once, clearing dialogs on autopilot and sending a half-written bug report I never meant to send.

One more worth flagging: mcp_tool hooks now support "onFailure": "block", matching the option command and HTTP hooks already had. If a hook call fails, the action it was guarding gets blocked instead of waving through on a technicality.

There's also a new warning for anyone running auto or bypass-permissions mode: a message from a session that skips permission prompts can now be held and dropped by a cloud session, and 2.1.328 tells you when that happens instead of letting it vanish silently. If you've ever wondered why a background agent seemed to go quiet mid-task, this is the kind of gap it's meant to close.

The self-hosted runner side of this release is heavier, new flags for labeling sessions and naming repository folders, but that's orchestration plumbing most solo setups never touch. Anthropic lists 91 bug fixes in this release altogether, most of them terminal rendering, paste handling and Code Mode edge cases nobody outside a specific setup will ever hit. The provider rollout and the two command changes above are the ones that'll actually show up in a normal session today.

What would I actually check today?

If I ran Claude Code against Bedrock, Google Cloud's Agent Platform or Microsoft Foundry for any client or employer work, the first thing I'd do is run /status and look at the permission mode line. Don't assume it still says what it said last week.

The changelog is blunt about this: no permission mode set means auto mode now, on three providers where that wasn't true a version ago.

Ten seconds with /status tells you which side of that line you're on.

What I wouldn't do is panic and set disableAutoMode everywhere out of habit. The classifier's block list, force pushes, credential leaks, unreviewed merges, production deploys, hasn't changed shape since August, and it's a reasonable one. I'd rather read what it actually blocks for ten minutes than switch it off because it's new to me.

I'd also go split the old /code-review habit into two commands now, today, rather than wait until a cleanup suggestion I actually needed quietly stopped showing up. Running /code-review out of muscle memory and assuming it still catches everything it used to is exactly the kind of small drift that's easy to miss for weeks.

For my own setup, raxxo.shop runs through the Anthropic API directly, not Bedrock or Foundry, so this particular change doesn't touch my permission mode at all. It would if I ever took on client work routed through someone else's cloud account, and that's exactly the scenario where I'd want to know the default changed before I found out the hard way, mid-session, watching something run that I expected to approve first.

The number that actually matters here isn't a percentage or a benchmark. It's three. Three cloud providers added auto mode in one release, on top of the two platforms it already covered since August. That's the whole scope of today's change, not a wider AI safety story, just a permission default catching up to where it already applied everywhere else.

Bottom Line

2.1.328 doesn't introduce a new idea, it closes a gap in one Anthropic already shipped. Auto mode was always going to reach Bedrock, Google Cloud's Agent Platform and Microsoft Foundry eventually, since leaving three major ways to run Claude Code on the older prompting model was never going to last. The thing to act on isn't the feature, it's the default. If you run sessions through any of those three and haven't set a permission mode on purpose, one is now set for you.

Check /status, decide if auto mode fits the work you do there, and use disableAutoMode if it doesn't. I wrote about why the default flipped in the first place back in August, and about the permission-prompt changes in 2.1.321 if you want the fuller history. Two releases ago, 2.1.327 fixed a worktree bug that had nothing to do with permissions at all, which is the usual pace here: one real change buried in a release full of smaller fixes. This one just happens to be the kind worth reading before your next session starts.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.