CHAPTERS
- 0:23 – 1:24
Why Claude Code updates matter: two themes (DX + autonomy)
Dixon Tsai frames the talk as a narrative behind recent rapid-fire releases, grouping them into two pillars: developer experience and autonomy. He sets expectations that the goal is better day-to-day usability while enabling Claude to handle more long-running work unattended.
- •Claude Code ships frequently; the talk explains how features fit together
- •Two themes: developer experience (DX) and autonomy
- •Emphasis on improving the human experience even as automation increases
- •Talk will end with ways to learn more beyond what’s covered
- 1:24 – 3:26
Remote Control: continue the same session from phone/web
Remote Control enables you to start a Claude Code session on your machine and then access that same terminal and context from Claude Code on the web (including mobile). The idea is to kick off long-running work, leave your desk, and still intervene or prompt the session as needed.
- •Start locally, pick up the same session remotely (web/mobile)
- •Remote view mirrors the terminal and dev setup in real time
- •Bidirectional input: type in web and see it in terminal (and vice versa)
- •Useful for long-running sessions with accumulated context across PRs
- •Enables spinning up sub-agents while away from keyboard
- 3:26 – 4:30
Flicker-free Terminal UI via full-screen mode + virtualized scrollback
The team addresses terminal flickering by moving to a full-screen TUI that virtualizes scrollback instead of appending in-place. This eliminates repaint artifacts, enables clickable UI elements, and keeps memory usage stable for long sessions.
- •Flickering came from misalignment/repainting in prior scrollback approach
- •New full-screen TUI virtualizes scrollback for smooth rendering
- •Clickable terminal UI elements become possible
- •Flat memory usage even for very long sessions
- •Team has largely switched to this mode internally
- 4:30 – 6:31
Full-screen TUI demo: voice mode, smooth output, and jump-to-bottom UX
A live demo shows enabling full-screen mode and using voice mode to issue prompts while Claude writes files. The chapter highlights smooth rendering, expandable tool calls, and scrolling behavior that doesn’t interrupt your place in the transcript.
- •Enable via “TUI full screen” (Terminal UI)
- •Voice mode toggle via /voice; hold space to speak prompts
- •Smooth rendering even for long content; no artifacts
- •Folded tool calls can be expanded with a click
- •Virtualized scrolling: “Jump to bottom” and new-message indicator
- 6:31 – 8:33
Revamped Desktop/GUI workflow: manage many sessions and richer context views
Dixon shifts to the Claude Code GUI (desktop/web), emphasizing improved session management and context visibility. The new UI supports organizing sessions and quickly switching between views like plan, diffs, and files to work more efficiently.
- •Sidebar updates: pinned sessions plus filtering/grouping (e.g., by project)
- •Drag-and-drop split view for parallel work
- •Multiple session views beyond transcript: plan, diff, files
- •Inline commenting and aggregation/resolution by Claude
- •UI features are optional “grab tools” when you need them
- 8:33 – 10:04
Experimental “Pin as chapter”: build a table of contents for long transcripts
An experimental feature lets you pin specific assistant messages as chapters, generating named anchors within a long session. This becomes a navigable table of contents that you can rename or delete to improve transcript navigation.
- •Hover assistant message → “Pin as chapter”
- •Auto-generates a chapter title; titles are editable
- •Creates a table of contents to jump around the transcript
- •Supports renaming and deleting chapters as needed
- •Positioned as an experiment with more desktop innovation to come
- 10:04 – 10:35
End-to-end GUI demo: run the app and verify trapezoid support
The demo returns to the original Excalidraw goal—adding trapezoids—then uses the integrated terminal to run a local server and validate the feature visually. It shows how Claude Code’s UI helps move from planning/editing to live verification quickly.
- •Open integrated terminal from the desktop session
- •Run the dev server and confirm localhost port
- •Validate the new trapezoid shape in the running app
- •Demonstrates quick iteration loop inside the GUI
- •Reinforces DX theme: less friction from idea to verification
- 10:35 – 11:05
Autonomy theme: reduce “why did you stop?” failures during long runs
Dixon motivates the autonomy features by describing common failure modes in unattended agent runs: permission prompts, conflicts, and missing commands. The next set of features aims to increase confidence that work continues correctly while you’re away.
- •Long-running tasks often stall on small blockers
- •Returning to a stuck run is frustrating and time-wasting
- •Goal: fewer interruptions and safer parallelism
- •Features are designed to increase trust during unattended execution
- 11:05 – 12:05
Auto Mode: fewer permission prompts with safety classification
Auto mode is a new permissions mode where Claude uses a classifier to decide whether to proceed with tool calls. It checks for destructiveness and prompt injection; safe actions run automatically while risky actions are blocked.
- •Eliminates repetitive permission approvals during agent work
- •Classifier evaluates: (1) destructive actions, (2) prompt injection signals
- •Safe tool calls proceed without user intervention
- •Unsafe calls are blocked and the agent must adapt
- •Reduces stalled long-running tasks caused by prompts
- 12:05 – 13:06
Worktrees: parallel features without agents stepping on each other
Worktrees make it easier to run multiple Claude sessions in parallel on separate branches/checkouts without collisions. Claude Code smooths Git worktree rough edges, adds shared-file configuration, and provides tools for Claude to enter/exit worktrees autonomously.
- •Run multiple feature efforts simultaneously in isolation
- •Claude Code integrates worktree setup and reduces Git’s sharp edges
- •Configure shared files across worktrees (e.g., node_modules)
- •Start sessions with --worktree (or -W) or ask Claude to create one
- •Claude gains enter/exit worktree tools; cleanup can be automated
- 13:06 – 16:12
Worktrees demo: two features in parallel + Claude-created star worktree
The demo launches separate worktrees for adding a new color and adding a border-radius slider, showing parallel progress. It also demonstrates Claude creating a new worktree for a star shape, then exiting and cleaning up when told to stop.
- •Kick off two independent sessions: color addition and border-radius slider
- •Demonstrates concurrent agents working on different PR-shaped tasks
- •Prompt Claude to create an additional worktree (star shape)
- •Claude asks when to implement; can exit and auto-clean worktree
- •Run and verify one completed worktree on a separate localhost port
- 16:12 – 17:43
Auto Memory: project knowledge persists as Claude-managed files
Auto memory reduces repeated setup by letting Claude store useful project facts over time. It maintains a per-project memory directory with a Claude-managed memory.md index that can reference more detailed files, and it extends to sub-agents as well.
- •Remembers build commands, debugging notes, and preferences across sessions
- •Claude decides what is worth remembering for future utility
- •Per-project memory directory with memory.md included in context
- •Progressive disclosure via links to additional memory files
- •Works for sub-agents (memory user/project) and benefits from newer models
- 17:43 – 19:14
Multi-agent Code Review: multiphase findings + verification (Ultrareview)
Anthropic’s code review workflow uses multiple reviewer agents focusing on different aspects, then verifies which findings are worth reporting. You can enable it via a GitHub app or run it manually in Claude Code with /ultrareview.
- •Pattern: multi-agent, multi-aspect review followed by verification
- •Catches issues that may take hours or be missed by a single pass
- •Available as a GitHub app for PR workflows
- •Manual entry point in Claude Code: /ultrareview
- •Complements other autonomy features to reduce human review burden
- 19:14 – 22:46
Routines + /loop: scheduled, connector-driven automation (local or remote)
Routines (research preview) let you configure prompts, repos, and connectors once, then trigger sessions via cron, GitHub webhooks, or API endpoints—optionally running remotely on Anthropic servers. The demo contrasts lightweight in-session /loop scheduling with a real routine that generates a daily GitHub issues digest and can post to Slack.
- •Routines run without manual prompting; triggers: cron, webhook, or API endpoint
- •Remote routines don’t require keeping your laptop open
- •/loop enables scheduling inside a single session (e.g., tell a joke every minute)
- •Example routine: daily scan of top-reacted GitHub issues filtered by labels
- •Connectors (e.g., Slack) can deliver digests; can chain to sub-agents for follow-up
- 22:46 – 24:55
Tool Search, enterprise improvements, and ways to keep up with releases
The talk closes with additional updates not covered in depth, including tool search (to expose more tools without bloating context) and enterprise/team features. Dixon points to the public changelog, social channels, documentation digests, and a newsletter for ongoing updates.
- •Stop loops by natural language; Claude maps intent to tool calls
- •Tool search adds indirection so more tools can be offered without context cost
- •Enterprise/admin improvements: better Windows support, cloud setup, native install, admin settings, plugin harness tooling
- •Follow updates via public changelog, Claude Devs on X, weekly docs digests, and developer newsletter
- •Closing note about upcoming on-stage conversation and thanks
