AI News
AI News AgentProductAnthropic3 min read

Claude Code gains autonomy with Sonnet 4.5

Anthropic has updated Claude Code with a beta VS Code extension, a redesigned terminal, and checkpoints for undoing changes. With Sonnet 4.5 as its default model, the assistant can take on longer development tasks without taking away your control over the code.

Claude Code can now handle longer, more complex development tasks with less constant supervision. Anthropic has added a VS Code extension, a redesigned terminal interface, and a checkpoint system for quickly undoing changes.

The assistant uses Claude Sonnet 4.5 as its default model and is now available to Claude Code users. The idea is not to let it code without limits, but to give you broader tasks to delegate while keeping the ability to review or restore your project.

Claude Code comes directly to VS Code

The new native VS Code extension is available in beta. It adds a dedicated sidebar where you can work with Claude and see its changes in real time through integrated code diffs.

This reduces the constant switching between the editor and the terminal. For example, you can ask Claude to prepare a new function, review its proposal line by line, and continue working from the same environment.

A terminal that is easier to control

Anthropic has also updated the Claude Code terminal interface. It now shows task status more clearly and lets you search your instruction history with Ctrl+r, which is useful when you want to reuse or modify an earlier request.

The company is also renaming its toolkit for building agents: the former Claude Code SDK is now called the Claude Agent SDK. It provides access to the tools, context management systems, and permissions used by Claude Code, along with support for subagents and hooks.

Subagents let you divide a job into specialized tasks. While one prepares a backend API, another can work on the interface. Hooks, meanwhile, trigger actions automatically at specific moments, such as running tests after modifying code or checking formatting before a commit.

Checkpoints for safer work

The main new feature for autonomous tasks is checkpoints. Claude Code automatically saves the state of the code before each change, so you can return to an earlier version by pressing Esc twice or using /rewind.

When you return to a checkpoint, you can choose what to restore:

  • The code.
  • The conversation with Claude.
  • Both elements.

The system only covers changes made by Claude. It does not restore the user's manual changes or commands run in Bash, so Anthropic recommends using it alongside a version control system such as Git.

What this changes for you

These features make it more practical to delegate tasks such as a large-scale refactor, exploring a new feature, or running lengthy processes. Background tasks can keep a development server running while Claude continues working on another part of the project.

The VS Code extension is in beta. The terminal improvements and checkpoints are available after updating the local installation, and the SDK lets teams build agents for specific areas such as financial compliance, cybersecurity, or code debugging.

The important change is not that Claude Code no longer needs review. It is that it can now work for longer with tools to inspect, automate, and undo its decisions. The question to watch is how far these features can go in real projects without increasing the number of errors a person must later fix.

Claude Code gains autonomy with Sonnet 4.5 | neversleep.ai