Last week at the London event, Ralph from Anthropic ran through what's shipped in Claude Code over the last few months. The talk was billed as a twenty-minute summary. It ran past thirty, and still left a pile of features uncovered.
The talk split into two buckets: things that change how it feels to use Claude Code, and things that let Claude do more without checking in with you. That split is a useful one. So I'll follow it.
Below is what shipped, what each piece is actually for, and which defaults I'd flip on without thinking.
Developer experience
Remote control
Start a session on your laptop, then drive it from the Claude mobile app or any browser. You get a push notification whenever Claude needs an input.
Anyone who's kicked off a long-running task and then realised they'd be away from the keyboard for an hour will recognise the value. You can be on the train, walking the dog, or sat in a meeting, and still drop a one-line prompt without going back to your desk.
There's a settings.json toggle to make every new session remote-controllable by default. The Anthropic team uses that as the standard. So do I, now.
Flicker-free TUI
This one looks minor on the changelog. It's the biggest quality-of-life improvement of the lot.
Claude Code's terminal had a long-standing problem: appending output to a large scrollback re-painted the whole terminal. On long sessions, the screen flickered every time Claude wrote a token.
The fix is a new full-screen mode. The terminal now virtualises the scrollback. Only the visible portion renders, memory usage stays flat, and the flicker is gone. As a side effect, the team can put UI affordances inside the terminal that weren't possible before: expandable file diffs, a jump-to-bottom indicator, clickable buttons.
Enable it with /TUI full-screen. Make it the default and don't look back.
Desktop app
The Claude Code desktop app got a full UI refresh. If you'd written it off after an earlier version, take another look.
The bits I'd actually use it for:
- Sessions grouped by repo. All your parallel sessions on a given project in one column.
- Plan-mode review with inline comments. Same idea as code review, but applied to the plan before any code is generated. Select a line, leave a note, send it back.
- GitHub diff inspection with line-level "explain this". Right-click a line in a diff, ask Claude what it's doing, get an answer that doesn't pollute the original session.
The desktop isn't meant to be your eight-hour-a-day environment. It's a second surface for the workflows the terminal is bad at. Specifically: visual review.
Autonomy
Now the substantive part. Every feature in this section has the same root cause: long-running Claude sessions used to stall, waiting on you, on something Claude could have decided itself.
Auto mode
The headline change. Instead of stopping at every permission prompt, Claude runs each action through a classifier with two questions:
- Is this destructive? Will I regret this later?
- Does this look like prompt injection?
If both answers are no, Claude proceeds. If either is yes, Claude looks for a non-destructive workaround first, and only stops to ask if there isn't one.
This is a philosophical move. The old default was "interrupt the human on every action," which made sense when sessions were short and supervised. With model capability where it is now, the friction was paying for nothing. Auto mode becomes the default for most flows.
A note: the classifier is a classifier. It will be wrong sometimes. Keep your hooks doing the truly non-negotiable enforcement: guards that don't run on probabilities.
Native worktrees
Worktrees aren't new to git. They've been usable with Claude since the beginning. They were a pain to manage, so most people didn't.
Now there's first-class support. A flag spins one up, and both the desktop app and the agents view create them transparently. Each parallel Claude gets its own segregated copy of the repo. They stop stepping on each other's files. The worktree gets cleaned up when the session ends.
If you run more than one Claude at a time, and after a few months of parallel sessions I can't picture going back, this is the feature that finally makes it pleasant.
Auto memory
Mental model: your CLAUDE.md is the onboarding doc you hand Claude on day one. Auto memory is the notebook Claude keeps while it works.
Claude watches the session, decides what's worth remembering (coding conventions, architectural choices, debugging patterns you've taught it), and writes those into a memory.md file. Critically, memory.md is mostly an index. The actual entries live in separate files. The index loads into context. The specific memories only load when relevant. No context bloat.
The files stay on your machine. They aren't pushed to GitHub. They're shared across every session and worktree in the same project. Audit them anytime with /memory.
This pairs naturally with the CLAUDE.md file. The manual stays for explicit rules. The memory grows for the implicit ones.
Code review (/ultrareview)
A practice the Anthropic team uses internally, now shipped as a feature. Every PR gets reviewed by a multi-agent pipeline:
- Phase one: a set of agents each look for a specific class of issue. Bugs. Vulnerabilities. Logic errors. Style violations.
- Phase two: a validator re-runs each finding against the actual code, throwing out false positives before they reach you.
Run it via the GitHub app (auto-triggered on PR), or invoke /ultrareview from your terminal. It's how Anthropic's own teams do PR review. That's enough of an endorsement for me.
Routines
This is the strategic announcement of the talk, even if it didn't get top billing.
A routine is a Claude Code session that runs on a schedule, an API call, or a webhook, without your computer being on. The session has the same capabilities as a regular one: file edits, tool calls, MCP connectors.
Examples: every time a GitHub issue is opened, triage it. Every morning at 08:00, check upstream for new releases of a dependency. Every sale your e-commerce backend records, run a routine against the inventory repo.
In research preview now. Worth watching closely. This is the surface where "Claude doing work" stops being a thing you launch and starts being a thing running in the background. Treat it accordingly: scope it tight, audit the secrets, log the runs.
Agents view
A dashboard for everyone running N Claude sessions at once. Sessions run in the background. The view groups them by state: working, needs input, complete. You can peek into any of them, or send a prompt to one without entering the session at all.
If you've been using tmux panes or six terminal windows for parallel sessions, this replaces them. Public preview.
What I'd flip on by default
Boring but high-impact:
- Remote control: always-on. Costs nothing. You'll use it within a week.
- TUI full-screen by default. Flicker is dead, leave it dead.
- Auto mode. Pair it with hooks and ADRs for the non-negotiables. See Stop Asking Your Agent Nicely.
- Worktrees by default for parallel sessions. If you're not parallel yet, you will be.
Things to try, with care:
- Auto memory. Audit
/memoryevery couple of weeks, especially after multi-day sessions. Delete what's wrong. /ultrareviewon real PRs. It costs tokens. Worth it on anything non-trivial.- Routines. Start with low-stakes triage. Don't point one at anything that pushes code until you trust the loop.
Where to follow along
The Claude Code changelog moves faster than any conference recap can. Subscribe to the changelog, follow the @claudedevs team on socials, and keep an eye on the "what's new" section in the docs. Or wait for the next event and someone, probably Ralph, will summarise three months of releases in twenty minutes again. And run over by ten.