On September 29, 2025, Anthropic announced updates to Claude Code aimed at letting the coding assistant handle longer tasks with less supervision. The release makes Sonnet 4.5 the default model and brings together checkpoints, specialized subagents, workflow hooks, and background tasks.
The practical problem is familiar: writing code is only part of development. A coding assistant also needs ways to recover from bad edits, separate unrelated work, run checks, and keep track of processes that take time. These features address those parts of the job, alongside a beta VS Code extension, terminal improvements, and an SDK for building custom agents.
Checkpoints make Claude's edits easier to undo
Claude Code now automatically saves the current state before making code changes. Pressing Escape twice or using /rewind allows a developer to restore code, conversation history, or both to an earlier checkpoint.
That distinction is useful during a refactor. Sometimes the code needs to go back while the discussion about the approach remains useful. In other cases, both the edits and the conversation need to return to an earlier point.
The recovery boundary matters. Anthropic says checkpoints apply to Claude's edits, not user edits or changes made through bash commands. A checkpoint therefore isn't a complete snapshot of everything that happened in the development environment.
This makes checkpoints a useful recovery tool alongside version control, rather than a replacement for it. They can reduce the cost of trying a larger change, but their scope limits what can be undone. A failed operation outside Claude's tracked edits still needs another recovery method.
Subagents separate work into focused tasks
Subagents let Claude Code delegate specialized work to agents that operate independently, each with its own context. Several parts of a project can move forward in parallel, with their results brought back into the main workflow.
A full-stack application illustrates the idea. One subagent could work on the backend API while another handles frontend components. Keeping their contexts separate gives each agent a narrower problem to work through instead of placing every detail in one conversation.
The potential benefit extends beyond speed. A developer can assign bounded pieces of work without managing every intermediate step in the same session. For a project with several moving parts, that could make delegation more practical.
Integration remains part of the work, though. Independent progress on an API and a frontend is useful only if the results fit together. Parallel execution changes how tasks are organized; it doesn't remove the need to check the combined result.
Hooks automate routine follow-up actions
Hooks trigger actions at specified points in a workflow. Instead of repeatedly asking for the same follow-up step, a developer can configure it to run automatically.
Running a test suite after code changes is a straightforward example. Code quality checks and documentation generation are other possible uses. Commit-related actions and deployment triggers are also potential areas for automation, depending on the available hook events and the surrounding setup.
The value is consistency. A routine check no longer depends on someone remembering to request it after every edit. This gives Claude Code a way to follow configured workflow rules, rather than relying entirely on conversational instructions each time.
Background tasks keep long processes from blocking a session
Builds, tests, deployments, and data processing can take long enough to interrupt other work. Background tasks allow those processes to keep running while the developer continues with something else.
Claude Code maintains awareness of the background operations and can report when they finish or encounter problems. A 10-minute build, for example, doesn't have to occupy the whole session while it runs.
This is especially useful alongside delegated work. Some tasks need active reasoning, while others mostly need time to execute. Keeping long-running processes in the background gives those activities room to proceed without losing track of their status.
More options in the editor and terminal
The beta VS Code extension brings Claude Code into the editor, with sidebar panels and inline diffs that show changes as they happen. Developers who prefer an IDE can review the assistant's work there instead of moving to a separate terminal interface.
The terminal experience also gets an update. Improved status visibility makes ongoing work easier to follow, and searchable prompt history is available through Ctrl+r. These are smaller changes than subagents or checkpoints, but they address frequent interactions in a tool used throughout the day.
For organizations building their own agent-based tools, the Claude Agent SDK, formerly the Claude Code SDK, provides building blocks for custom experiences. Possible applications include development tools tailored to a team's domain, workflow, standards, or infrastructure.
Sonnet 4.5 becomes the default
The release makes Sonnet 4.5 the default model in Claude Code. Users can still choose another model with the /model command.
Anthropic positions Sonnet 4.5 for extended, autonomous reasoning. Longer development tasks require the model to retain relevant context, assess the effects of code changes, and coordinate delegated work. Those demands go beyond answering an isolated coding question.
The model and the surrounding tools serve different roles. The model reasons about the work, while features such as automatic checkpoints and hooks provide recovery and execution mechanisms around it. Their usefulness depends on how well those pieces work together during a real task.
What the release makes possible
The update gives developers more ways to delegate a sequence of work rather than request one answer or edit at a time. Subagents can divide tasks, hooks can run follow-up checks, background processes can continue without blocking the session, and checkpoints can undo tracked edits when an approach fails.
For individual developers, this could make larger refactors and multi-part changes easier to manage. Teams could use the same capabilities for AI-assisted code review, automated testing, and documentation generation. These are possible workflows, not evidence that the release guarantees better code or less review work.
As of October 1, 2025, the announced capabilities are available to Claude Code users, with the VS Code extension in beta. The extension is available through the Extension Marketplace. The terminal improvements and checkpoints require updating the local installation, while developers building custom agents can consult the Claude Agent SDK documentation.
Anthropic's official announcement provides further details on the release.
