Anthropic announced on September 29, 2025 a series of improvements to Claude Code that make it more capable of working autonomously inside the terminal and in IDEs. What does this mean for people who code and for teams that want to delegate complex tasks to an agent? I explain it with concrete examples and without unnecessary jargon. (anthropic.com)
Qué cambia en Claude Code
The key update is that Claude Code now runs Sonnet 4.5 as the default model, which gives it more memory and the ability to handle long, complex tasks. In addition, Anthropic adds three important blocks: a native VS Code extension, a refreshed terminal interface, and a checkpoint system to delegate safely. (anthropic.com)
Nuevas superficies y experiencia de desarrollo
Extensión nativa para VS Code
If you prefer working in an IDE, there’s now a beta extension for VS Code that integrates Claude Code into a sidebar. You’ll see real-time changes with inline diffs, which makes it easy to review and accept edits without leaving your editor. You can download it from the marketplace if you want to try it now. (anthropic.com)
Interfaz de terminal mejorada
Claude Code’s terminal got a refresh: better state visibility and a searchable prompt history with Ctrl+r. That’s useful if you reuse or tweak previous prompts during long work sessions. (anthropic.com)
Autonomía controlada: checkpoints, subagents y hooks
Anthropic introduces a checkpoint system designed to give confidence when delegating long tasks. Before each change, Claude saves the code state. If something goes wrong, you can go back by pressing Esc twice or using the /rewind command. Important: checkpoints apply to Claude’s edits, not to your direct changes or bash commands, and Anthropic recommends using them alongside version control. (anthropic.com)
Use checkpoints when you try wide refactors or design explorations. They’re a safety net, not a replacement for version control.
In addition to checkpoints, Claude Code now supports subagents, hooks, and background tasks:
- Subagents: delegate specialized work in parallel, for example spinning up a backend while the main agent works on the frontend.
- Hooks: trigger automatic actions, like running tests after changes or applying lint before commits.
- Background tasks: keep dev servers or long processes alive without blocking the agent’s progress.
These pieces together let you assign broader-scope tasks to Claude Code, but with mechanisms to regain control when needed. (anthropic.com)
Claude Agent SDK: personaliza agentes para tu flujo
The Claude Agent SDK (formerly Claude Code SDK) exposes the tools, context handling, and permission framework that Claude Code uses. There’s now support for subagents and hooks in the SDK, which makes it easier to build custom agents for needs like financial compliance, security, or automated debugging. If you’re building enterprise flows, this is where you configure permissions and behaviors. (anthropic.com)
Ejemplos prácticos para entenderlo mejor
-
Startup that needs to ship a feature: the main agent builds the UI while a subagent spins up the API and another runs tests in parallel. Result: faster iteration without one process blocking the others.
-
Security team: a specialized agent performs static analysis, another runs fuzzing tests, and hooks generate automatic reports when vulnerabilities are found.
-
Massive refactor: you let Claude explore and propose changes; if something doesn’t fit,
Esctwice and you go back to the previous checkpoint without losing the rest of the work.
These cases are already being explored by developers using the SDK and the new capabilities. (anthropic.com)
Cómo empezar hoy
- Sonnet 4.5 is the default model in Claude Code; use
/modelif you need to switch models. (anthropic.com) - The VS Code extension is in beta and can be downloaded from the Marketplace if you want to try the integrated experience. (anthropic.com)
- Terminal updates and checkpoints are available to Claude Code users: update your local installation to access them. (anthropic.com)
- For developers who want to build agents, the Claude Agent SDK docs show how to get started and configure subagents and hooks. (anthropic.com)
Qué significa esto para desarrolladores y equipos
Does AI replace the programmer? Not exactly. These improvements are meant to speed up repetitive tasks, orchestrate parallel work, and reduce friction in broad explorations. You still decide the limits, review changes, and set permissions.
If you’re responsible for product or engineering, the practical advice is to integrate checkpoints and version control from day one, define agent permissions, and start with experimental flows in staging environments. That way you get autonomy without losing control.
Claude Code is moving toward more useful, controlled autonomy: more surfaces to use AI, tools to customize behavior, and safeguards so delegating isn’t a leap into the void. If you want, I can help turn one of your workflows into a small plan to try these features.
