Live Artifacts, Auto Mode Guardrails, and a Jumper Hire

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of June twenty-second. Claude Code sessions can now produce live, interactive artifacts. Auto mode has new guardrails for destructive commands. And John Jumper is leaving Google DeepMind for Anthropic.

Let's start with Claude Code -- the headline this week is artifacts, which let you preview in-progress work from a session as a live, interactive thing your teammates can actually click on. The artifact is built from your full session context, stays linked to that session, and updates as the work progresses. That means no exporting code, no spinning up a separate dev server just to show someone what you've built. The artifacts announcement on the Anthropic blog has the full picture, and we'll link it in the show notes.

The same theme of removing friction shows up in agent teams, which got a meaningful simplification. If you're running the experimental agent teams flag, the separate tools for creating and deleting teams are gone. Every session now has one implicit team, and you spawn teammates directly through the agent tool. Existing prompts that reference team names still work -- the parameter is just ignored now -- so nothing breaks.

While teams got simpler, auto mode got more careful. Auto mode now blocks destructive git commands when you didn't ask to discard local work -- things like hard resets, force-clean, and stash drops. It also blocks amending commits the agent didn't make in the current session, and it blocks terraform destroy, pulumi destroy, and cdk destroy unless you explicitly named the stack. Subagent spawns are now evaluated by the same classifier before launch, so the guardrails extend down the agent tree.

On the configuration side, you can now change any setting inline from a prompt -- one command, key equals value, and you're done. It works in interactive sessions, in headless mode, and through Remote Control. And permission rules got more precise: you can now match on tool parameters with wildcards, so a rule can target, say, only Opus subagents instead of blocking the agent tool wholesale. One more workflow note -- when you have nested project directories with their own Claude configuration, the one closest to your working directory now wins on name collisions, and nested skills load automatically when you're working in that subtree.

Moving under the hood, the theme this week is fewer things getting lost. Viewing a subagent's transcript now shows tool results and live progress as it happens. Messages you send while a subagent is finishing its turn are no longer dropped on the floor. Backgrounding a running subagent with Control-B no longer restarts it from scratch.

And if your network connection drops mid-stream, partial responses are preserved instead of disappearing behind a raw error. Background tasks started by a teammate also keep running after that teammate finishes its turn, instead of getting killed. On the performance side, a regression that was adding about one hundred twenty milliseconds to every launch in fresh environments has been fixed. Startup also no longer blocks with a blank terminal for up to fifteen seconds when the account settings fetch is slow.

And if you work off network drives or cloud-synced folders, write and edit operations no longer occasionally produce zero-byte or truncated files -- which is the kind of bug you really want gone.

Over to the broader Anthropic news, and the policy story from last week is still developing. The Wall Street Journal is reporting that Anthropic and the Trump administration are in active talks to restore access to Fable five and Mythos, following the restrictions that landed last week. If your team's model routing has been affected, that's the one to watch. Wired has a companion piece digging into the technical demand at the heart of the standoff -- specifically, that Anthropic guarantee no jailbreaks of its frontier models -- and why researchers say that's not actually achievable.

Useful context if you want to understand what the policy fight is really about. In hiring news, CNBC is reporting that John Jumper, the Nobel laureate and AlphaFold lead, is leaving Google DeepMind to join Anthropic. That's a notable signal about where frontier research talent is heading, and possibly about what areas Anthropic plans to invest in next. And finally, the Wall Street Journal is also reporting on a new lawsuit targeting the usage caps on Anthropic's two-hundred-dollar-a-month Max plan -- the plan a lot of Claude Code power users are on.

Worth tracking if your daily workflow depends on those quotas. That's it for the brief. I'm Shannon, and we'll see you next week.