Guided Auto Mode Setup Reaches Every Backend

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of October twenty-sixth. Guided auto mode setup now works on every backend. Code Review gains depth levels. And Anthropic's IPO filing is public, with compute costs front and center.

Let's start with Claude Code, where guided auto mode setup now reaches every backend -- including the ones that previously had no path to it at all. Here's what it does: the setup walks through your project and your recent sessions to learn what your environment actually looks like, so auto mode stops asking you about the commands you run all the time. Until now, that only worked in certain configurations. It now runs on Bedrock, on Vertex, on Foundry, behind the Claude apps gateway, and in sessions with telemetry disabled.

Auto mode also got clearer about one of its prompts. The first time it wants to read a file outside your working directories, it now explains that blocking that read means your shell commands will start asking for permission instead. Small wording change, but it stops you from making a choice you didn't understand. Staying on the theme of configuration, Code Review picked up depth levels.

Organization owners can now set Quick, Standard, or Deep as the review level across the whole organization, then override it on a per-repository basis. You can also opt in every repository in a GitHub organization at once, and that includes repositories added later. And if you've got a standing review subscription running on a pull request, there's now a way to end it without triggering one more review on your way out. Now, the change most likely to catch you off guard involves in-place stream edits.

Running sed with the in-place flag no longer goes through the simulated file-edit preview. It gets the standard Bash permission prompt and runs exactly as written. Worth knowing: like other Bash edits, those changes are not restored by rewind. Over in Visual Studio Code, background subagents got quieter.

Instead of streaming every step into your chat panel, they now show as a single status line under their Agent card, with the full transcript still available in your tasks view. Edit and Write in the panel also got faster -- the wait for new problems after each edit dropped to three quarters of a second, down from one to one and a half. Two more. If you're on Pro or Max and you resume an inactive conversation that would use a noticeable share of your five-hour limit, Claude Code now tells you before you spend it.

You can resume anyway, or start fresh and let Claude look up the previous conversation if you refer to it. And usage warnings are now one actionable line each, naming where the problem came from, with a single keypress to apply the fixes. The same suggestions surface in Visual Studio Code and in the doctor check.

Under the hood, this was a week about closing gaps in the sandbox and permission model. Sandboxed Bash commands no longer run your shell startup files outside the sandbox. Deny-write patterns now apply through symlinks. Sandboxed and auto-accepted edits can no longer modify the file you passed in as your MCP configuration.

And Bash and PowerShell permission checks now prompt on compound directory-change commands and redirects that they previously waved through. On the same hardening note, ultrareview now requires you to confirm workspace trust in a terminal, and refuses to run anywhere else without it. There's a flag and an environment variable for scripted use. Separately, compaction got better behaved.

Sessions no longer compact over and over after reading a very large nested memory file, compaction after a prompt-too-long error now trims dense content like base64 and SVG properly, and non-interactive runs use less memory when compacting very large transcripts.

On to the broader Anthropic news, and it's mostly about money this week. The Wall Street Journal has an exclusive on Anthropic's S-1 filing, which lays out rapid revenue growth alongside compute costs that drove a large first-half loss. Those economics shape the pricing and capacity decisions behind the tools you use every day. Bloomberg follows that thread further, reporting long-term compute commitments of up to seven hundred fifty billion dollars, with a separate forty-five billion dollar deal with Nscale in talks.

Buildouts of that size are what sit behind model availability and rate limits. On the product side, the New York Times reports that Anthropic previewed Claude Duet, a personal agent app aimed at everyday tasks rather than coding -- a signal of where the agent patterns you already use in Claude Code are heading. And the Guardian has a piece on self-replicating AI worms and the risks Anthropic has been flagging publicly. It's useful context for the sandbox and permission hardening we just covered.

We'll link all four in the show notes. That's it for the brief. I'm Shannon, and we'll see you next week.