Claude Code 2.0: An Impressive Upgrade

Claude Code has been moving fast. A string of updates over the past month has transformed it from a capable coding tool into something closer to a genuine multi-agent development environment. Sessions can now talk to each other. The tool now handles dangerous commands more reliably than most human reviewers do. And a built-in browser and iOS simulator have quietly made the desktop app into a complete development workspace.

Here is everything that changed, in plain terms.

Sessions Can Now Talk to Each Other

The most significant new capability in Claude Code is cross-session messaging, which shipped in version 2.1.224. Before this, if you had two Claude Code sessions running simultaneously, they had no way to communicate. One session working on your database schema had no idea what the other session working on your API layer was doing. You were the relay. Every time something changed in one session that affected the other, you had to copy the information yourself and re-explain the context.

That is gone now. Sessions can discover each other using a list_agents command and send direct messages to each other using send_message. The messages are concise text summaries of the relevant information, not dumps of entire conversation histories or file contents. The receiving session gets exactly what it needs to stay synchronised.

The practical use is straightforward. Say you have one session working on your front end and another on your payments API. If a field name changes in the front-end session, you can tell Claude to let the payments API session know. Or Claude can do it automatically if it detects the change affects the other session’s work.

More broadly, this enables multi-agent orchestration that was previously impossible without external infrastructure. You can give sessions named roles and have them coordinate around shared state. A research session can gather information while a separate implementation session pulls from it without requiring you to transfer anything manually. Named sessions persist across runs, allowing a specialist session to be called repeatedly as a reference point.

Cross-session messaging currently works on macOS and Linux. There is no native Windows support yet. The feature requires Claude Code v2.1.224 or later.

Auto Mode Is Now the Default, and the Reason Is Surprising

Starting August 14, auto mode became the default permission mode for new Claude Code sessions on Pro, Max, and Team plans. If you have already set a custom default, it stays as-is. For everyone else, sessions now run autonomously rather than pausing before every file edit, shell command, or network request.

The reason for the switch is a study Anthropic ran with 1,053 paid developers. They found that humans caught only 13.6 per cent of dangerous commands introduced into a session, whereas the auto-mode classifier caught 89 per cent of the same commands. Human performance also got worse as sessions grew longer, a predictable result of approval fatigue. Theauto-modee classifier’s performance remained relatively consistent.

To understand what auto mode actually does, it helps to know what it is not. It is not the same as bypass mode, which skips all safety checks entirely. Auto mode works by running a separate classifier model (Sonnet 5 by default) that reviews each of Claude’s intended actions against your request and the conversation context before the action executes. If the classifier determines that an action is risky, irreversible, or appears to be driven by content Claude picked up from an external source rather than from you, it blocks the action and asks. If it looks fine, it runs.

In a separate third-party prompt-injection evaluation, Opus 5, Sonnet 5, and the Claude 3 model running in auto mode completed 720 adversarial attempts, with zero successful attacks.

The data Anthropic cited for the shipping decision includes teams using auto-mode shipping roughly 25 per cent more pull requests, largely because Claude can run uninterrupted for longer sessions. Enterprise and API access remain opt-in for now.

You can switch modes at any time. In the CLI, press Shift+Tab. In the desktop app, use the mode dropdown.

Concise Mode Saves Tokens Without Losing Detail

Claude Code now has a concise output setting that tells Claude to lead with results and cut unnecessary explanation, while still providing full detail whenever you ask for it. You enable it through the config file.

The practical value is in long coding sessions. Claude Code’s explanations can be thorough to the point of being verbose, and verbose responses eat through your weekly usage limit faster than the actual work does. Concise mode shifts the default toward action over explanation without removing the option to ask for more context when you need it.

It is a small setting with a meaningful effect on how far your weekly limits actually stretch.

A Built-In Browser and iOS Simulator

Two additions have quietly made Claude Code’s desktop app into a more complete development environment.

The desktop app now has a built-in browser. Claude can open documentation, design files, dashboards, or any website directly inside the session, read through pages, click around, and use what it finds while it works on your code. Instead of copying information from a browser window into your terminal, Claude fetches the context itself. The browser is sandboxed and configurable. You control whether sessions can log into things with your credentials, and persistent login can be enabled through session settings.

The iOS simulator integration is more specific but genuinely useful for mobile developers. On macOS, Claude Code can now build and launch your iOS app, and the simulator opens alongside the conversation, just like artefacts do. Claude can tap through screens, verify changes, and interact directly with the live app. The development loop (build, test, debug, iterate) now occurs in a single window instead of switching between Claude Code and a separate Xcode simulator. This feature is in public beta for Pro, Max, and Team plans, and requires Xcode with the iOS platform installed.

Code Review Got Effort Levels and Ultra Mode

The slash code review command now supports adjustable effort levels, which changes how thoroughly Claude examines your code and at what cost.

At low effort, it runs a fast single-pass review over a diff. Cheap enough to run before every commit. Medium looks at the changed code in context, runs multiple passes from different angles, and verifies findings before surfacing them. High adds fresh context sub-agents as both finders and verifiers. Extra high looks for downstream issues outside the code you directly changed, the kind of secondary effects that only matter after a change lands in production.

Ultra mode spins up a fleet of independent reviewers and has them reproduce findings to cut false positives. Anthropic says they run this on every pull request internally.

On a human-labelled code review dataset, Claude Opus 4.8, using slash code review, finds about 17 per cent of bugs at low effort and roughly 25 per cent at extra high effort. The leading competitor increases from around 10 per cent to over 20 per cent over the same range. The trade-off is between speed and cost and depth of recall. Artefacts Can Now Connect to Live Data

A quieter but meaningful update: Claude Code artefacts can now connect directly to MCP servers. That means an artefact such as a dashboard, internal tool, or interactive app can fetch live information rather than displaying static data. When a viewer interacts with the artefact, it can take actions on their behalf through the connected service.

This turns artefacts into functional applications rather than generated interfaces. MCP access is restricted for publicly shared artefacts, so this applies to artefacts used within your own account or organisation.

Smaller Updates Worth Knowing

A few other additions round out the update.

The desktop app now launches roughly twice as fast as it did a month ago, after Anthropic fixed some background startup throttling. The CLI also uses about half as much GPU as it did at the 99th percentile, due to a change that prevents Bun’s garbage collector from running during active sessions.

The design skill, currently in research preview, adds a /design command that generates editable UI artboards inside your session. You choose a design, tweak it, and Claude implements it directly in your application.

A screen reader mode is now available for users who rely on VoiceOver or similar tools. It replaces the visual terminal interface with plain linear text, adds clear speaker labels, and can play a bell when Claude needs your attention.

Auto-continue is now available in the desktop app. When you hit your weekly usage limit, you can enable a checkbox at the bottom of the session. When your limit resets, Claude automatically resumes from where it left off without requiring you to restart.

Visual feedback is also available through the pencil icon in the desktop app. You can draw directly on a screenshot, highlight sections, add text, or drop in shapes to indicate where you want changes, and Claude uses that visual context as part of its next action.

The sum of all of this is a coding tool that is increasingly less about Claude answering questions and more about Claude doing work. Sessions that coordinate with each other, a classifier that outperforms human reviewers at catching dangerous actions, a browser and simulator built directly into the workspace. Each update is incremental. Together, they add up to something that is starting to look quite different from what Claude Code was six months ago.