anthropics/claude-agent-sdk-typescript
Claude Agent Sdk Typescript
该项目是一个 TypeScript 软件开发工具包，旨在构建和编排与代码库和系统环境交互的自主代理。它提供了一个程序化接口，用于构建能够执行复杂工作流的代理，例如通过利用大语言模型进行自动化代码重构、文件系统操作和 Shell 命令执行。
该框架通过对安全、受控代理操作的关注脱颖而出。它包括细粒度的访问控制系统，允许开发者为工具和外部服务定义特定权限，确保自动化任务保持在定义的安全策略内。该架构支持分层委托，使主代理能够为专注的任务生成和协调专门的子代理，同时保持有状态的会话持久性以允许恢复之前的工作。
该 SDK 提供对代理生命周期的全面可观测性和控制。开发者可以利用事件驱动的钩子在运行时拦截、验证或修改代理行为，并将实时执行进度和任务通知流式传输到宿主应用程序。该环境专为一致性和安全性而设计，具有隔离的执行上下文和可配置的系统指令，这些指令管理代理如何解释和作用于其环境。
Features
- Agentic LLM Frameworks - Provides a TypeScript SDK for building and orchestrating autonomous agents that interact with codebases and system environments using large language models.
- Autonomous Agents - Provides a framework for building intelligent agents that interpret codebases and execute complex workflows using large language models.
- Agent Governance Platforms - Provides granular access control systems to ensure automated agents operate safely within sensitive development environments.
- Agent Construction Frameworks - Provides a programmatic interface for constructing autonomous agents that interpret codebases and execute complex workflows.
- Agentic Task Orchestration - Coordinates multi-agent systems and manages complex development workflows through a programmatic interface for tool execution and task planning.
- Agentic Workflow Orchestration - Coordinates autonomous agents to perform complex tasks like codebase analysis and file editing through a structured interface.
- Hierarchical Task Delegation - Enables primary agents to manage complex workflows by spawning and coordinating specialized subagents.
- Agent Configurations - Allows defining custom system instructions to govern how agents interpret and act upon their environment.
- Parent Agent Sub-Agent Delegation - Supports hierarchical delegation by allowing primary agents to spawn and coordinate specialized subagents.
- Agent Tool Integrations - Allows linking external tool servers to grant agents direct access to specialized functions beyond built-in capabilities.
- Autonomous Agent Orchestration - Provides a toolkit for hooking into agent events, persisting sessions, and isolating execution environments.
- Agent Tool Integrations - Maps natural language requests to structured programmatic tool invocations for external service interaction.
- Shell Command Execution - Allows triggering shell commands and scripts from within an agent workflow to interact with the local environment.
- Programmatic Code Refactoring - Uses programmatic agents to read, modify, and update source code files for complex refactoring tasks.
- Agent Action Restrictions - Provides granular access control to restrict which tools an agent can use, ensuring sensitive actions are blocked or require explicit approval.
- Agentic Session Persistence - Stores conversation history and agent context in a persistent format to allow resumption of previous work.
- Agent Sandboxing Policies - Restricts agent interactions with the filesystem and external tools to ensure secure execution within isolated environments.
- AI Agent Permissions - Enables granular definition of access policies for tools and external services to ensure automated tasks remain within defined security boundaries.
- Codebase File Manipulation - Enables programmatic reading, writing, and modification of source code files to perform automated refactoring within a codebase.
- AI Agent Execution Monitors - Streams real-time execution progress, task notifications, and usage metrics to provide visibility into agent activity during runtime.
