Anthropic is turbocharging its AI coding assistant, Claude Code, with new features that make it more autonomous and easier to use for software developers. In a product update this week, the company unveiled a native Visual Studio Code extension, a revamped terminal interface, and an automated “checkpoint” system for Claude Code.
Powered by Anthropic’s latest Claude Sonnet 4.5 model, these enhancements aim to let Claude handle longer, more complex programming tasks – potentially even driving entire coding projects – while giving developers tools to supervise and intervene when needed.
Coding alongside you in VS Code: One major leap is the introduction of a Visual Studio Code extension (in beta) that embeds Claude Code into VS Code, the popular code editor used by millions of developers.
Previously, developers interacted with Claude via a separate terminal or web interface, copying code back and forth. Now, with the extension, Claude sits in a sidebar within VS Code, and developers can chat and receive suggestions contextually as they edit files.
The extension highlights Claude’s proposed changes with inline diffs – so if Claude suggests refactoring a function, you’ll see the exact lines it would change in your code editor. This provides a much more transparent and interactive experience. You can accept the changes, modify them, or ask Claude to regenerate.
It’s akin to having an AI pair programmer “live” in your IDE, observing the codebase and offering help in real time. Early testers note this also helps catch misunderstandings quickly – if Claude’s suggestion isn’t quite right, the developer sees it immediately and can correct or refine the prompt.
Better CLI experience: For those who prefer working in the command-line, Claude Code’s terminal interface version 2.0 comes with quality-of-life improvements.
Users will see clearer status indicators (so you know when Claude is “thinking” vs ready), and a searchable prompt history (via Ctrl+R) that lets you quickly pull up previous instructions or code generations to reuse or edit. This addresses a pain point where long interactive sessions would scroll off screen.
Now you can scroll or search through what you and Claude have been doing. Combined with the existing features like code execution within the CLI, the terminal mode is increasingly robust for those integrating Claude into their shell workflows or automations.
Claude as an autonomous coder – safely: Perhaps the most impactful update is the checkpointing feature.
As developers experiment with giving Claude more autonomy (like “Here, Claude, implement a new sorting algorithm across the codebase”), there’s understandable anxiety: what if it introduces bugs or deletes something important? Checkpointing directly tackles that by automatically saving the project state before every Claude-driven change.
If Claude is working in a repository, it will snapshot the files it’s about to modify. If at any point the results are not satisfactory, the developer can instantly rewind to a previous checkpoint by a simple command (double-tapping Esc or using /rewind).
And this isn’t just one level of undo – checkpoints are kept for each change, so you could go back multiple steps if needed. Critically, you can choose to restore just the code to the prior state, just the conversation context, or both.
This flexibility is important: maybe you want to undo the code changes but still remember the discussion you had with Claude leading up to it (so you can try a different approach without starting from scratch).
Or conversely, you might want to reset the conversation (if it got off track) but keep the code edits. By distinguishing these, Anthropic shows it understands the iterative nature of coding with AI.
They do caution that checkpoints cover Claude’s edits, not human edits or external commands, so developers should still use version control for their own changes. But checkpoints essentially encourage developers to let Claude try bold, wide-ranging modifications: you can always roll back.
Checkpoints become especially powerful when combined with Claude’s new autonomous features like subagents, hooks, and background tasks (features that were introduced in the Claude Agent SDK and integrated into Claude Code).
For instance, imagine you tell Claude to “Optimize our application’s performance.” Claude might spawn subagents: one to profile the code, another to refactor slow functions, another to adjust database queries – all in parallel. With hooks, after each refactor, it could automatically run your test suite to see if everything still passes.
And with background tasks, it could start a dev server or lengthy compilation in the background while continuing to work on something else. These are incredible capabilities, but they carry risk: multiple things changing at once could be chaotic.
Checkpoints ensure that even in this autonomous dance, there’s a safety net under every step. If one subagent’s change breaks something, you can rewind that part without losing the progress in others.
Building custom AI devs: Anthropic’s push with the Claude Agent SDK (formerly Claude Code SDK) is to let companies build their own specialized coding agents. Now supported in Claude Code is the notion of subagents and hooks, which advanced teams can script.
For example, a cybersecurity team might create a “vulnerability patching agent” that, when triggered, uses Claude to scan for a known security flaw pattern in code (subagent 1), automatically patch those spots (subagent 2), then run security test cases (hook after each patch).
The SDK offers these as templates and the permissions framework to ensure agents don’t exceed allowed actions. Several enterprises were already experimenting: we’re told developers have built agents for financial compliance (ensuring code changes meet regulatory requirements) and debugging (automatically finding and fixing common bugs).
The success of such agents will depend on both the power of the underlying model (Sonnet 4.5 is clearly up to many tasks) and the reliability of these guardrails. Anthropic’s strategy seems to be: provide the building blocks (Claude’s intelligence + tools + safety checks) and let domain experts craft AI solutions for their domain.
In that sense, the Claude Code update is not just about coding – it’s about enabling AI co-workers in any knowledge work that can be structured similarly to coding (like writing legal contracts, updating spreadsheets, etc., though coding is a clear starting point).
Competitive landscape: Microsoft’s GitHub Copilot is currently the most famous AI coding assistant, and it’s also adding features like a VS Code chat sidebar and, soon, its own voice-activated Copilot X and terminal integration.
Anthropic’s Claude Code, especially with Sonnet 4.5’s stronger performance, is a strong challenger. Some developers prefer Claude for its larger context window and generally more verbose explanations, which help in understanding why it’s suggesting something.
With these new features, Claude Code matches or exceeds Copilot in many areas: Copilot doesn’t yet have the concept of subagents or automated multi-step tasks, for instance.
OpenAI’s upcoming “assistant with tools” in the coding domain might respond, but Anthropic has taken a lead on the feature-rich approach to AI coding.
For developers, the bottom line is that working with Claude Code is becoming more like working with a knowledgeable teammate and less like using a glorified autocomplete.
You can converse with it in natural language inside your coding environment, it can proactively take on tasks (and you can trust it not to wreck your project thanks to checkpoints), and you can even customize its workflow with your own automations.
One early user described it as “giving an intern the ability to write code while I supervise with a finger on the undo trigger”.
That feels apt – Claude can now bravely attempt big refactors or feature builds, while you oversee and mentor it. Over time, as confidence in these systems grows, one can imagine even less oversight will be needed for routine tasks.
Anthropic’s latest update to Claude Code shows a clear vision: an AI that doesn’t just assist in coding, but can gradually take on the laborious parts of development, from writing boilerplate to running tests, allowing human developers to focus on higher-level design and critical thinking.
It’s a vision shared by others in the AI field, but Anthropic is executing rapidly to make Claude a compelling choice for those living on the cutting edge of AI-assisted programming.