Auto Mode Becomes the Default, Self-Hosted Runners Arrive

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of August tenth. Auto mode is now the default permission on Pro, Max, and Team plans. Self-hosted runners are in public beta. And Meta released a coding agent of its own.

Let's start with Claude Code -- and the change most of you will feel first is that auto mode is now the default permission mode on Pro, Max, and Team plans. That means the per-tool prompt flow you've been living with is gone. In its place, a classifier decides which routine actions get approved automatically and which ones are risky enough to stop and ask about. Practically speaking, you'll notice Claude running longer stretches on its own before pausing for input, and Anthropic says the classifier catches more destructive commands than the old allowlist did.

There's a companion post on running auto mode in production, and we'll link both in the show notes. That shift toward keeping work inside a controlled boundary shows up in the next release too. A new self-hosted runner, in public beta on Team and Enterprise plans, turns your own machines or containers into execution environments for Claude Code sessions started from the web, mobile, or desktop clients. Your code and your network access stay inside your infrastructure -- the client just orchestrates.

If you've been waiting to try the mobile or web clients but couldn't get past the network boundary, this is the piece that was missing. Sessions can also now message each other across machines. One Claude Code session can address another by name, on macOS and Linux, and a follow-up release lets you initiate a conversation with a remote session by name rather than waiting for it to reach out first. On the workflow side, the review command is now an alias for code-review, which reviews your current diff or a specific pull request.

You can ask for a deeper cloud review by passing a level, and if you leave the level off, it reuses whatever you typed last. And if you work in Visual Studio Code, there's a new Focus view that collapses tool activity behind an expandable per-turn summary, with a live indicator showing what's currently running. It's a keyboard shortcut away, or you can toggle it from the command palette.

Moving under the hood -- the theme this week is closing gaps around the permission system, which matters more now that auto mode is doing the approving. Several permission-check bypasses got patched. Zsh could hide commands inside regex conditionals. Crafted Bash commands could obscure parts of themselves from the checker.

Tab characters and invisible Unicode padding could hide text from the approval dialog. Workflow scripts could escape the sandbox through dynamic imports. And deny-read and deny-write entries with a trailing slash character were silently bypassable on Linux and macOS. Restart Claude Code to pick up the patches.

There's also a smaller but useful change on the plugin side -- you can now install plugins directly from an HTTPS zip URL, with optional hash pinning, so locked-down machines that don't have git or a package manager available can still pull plugins in. Forked sessions now get their own git worktree instead of sharing the parent's checkout, which means forked work can't stomp on the original branch anymore. And auto mode itself got a performance pass -- it now reuses the cached conversation prefix across parallel permission checks, which cuts cache costs on long sessions. The Stats panel now breaks down cache tokens by category too.

On to the broader news. The Wall Street Journal is reporting that Meta has released a coding agent of its own, putting it in direct competition with Claude Code and Codex. It's a signal that the landscape around the tool you use every day is getting more crowded. Related to that -- The Information has a piece on how firms like Coinbase are building their own coding agents on top of Claude Code rather than replacing it.

Useful context if you're weighing where custom infrastructure pays off around a stock install. On the infrastructure side, Bloomberg is reporting that Anthropic signed a ten billion dollar compute deal with a cloud startup called Volta, adding to a growing stack of capacity commitments behind the Claude models. And Reuters says Anthropic is standing up its own in-house chip design team and hiring engineers for it -- a longer-horizon bet on controlling more of the hardware stack behind Claude. We'll link all four stories in the show notes.

That's it for the brief. I'm Shannon, and we'll see you next week.