A Managed Deep Agents project has a required agent entry and optional files that enable managed capabilities. It is a regular TypeScript project. 托管深度代理目前处于私有测试阶段，仅在LangSmith Cloud的美国区域提供。加入等待列表以申请访问权限。

​
Project layout

唯一的必需文件是代理入口：agent.py、agent.ts 或 agent.tsx。它必须导出一个使用 define_deep_agent 或 defineDeepAgent 创建的、名为 agent 的定义。tools/ 和 middleware/ 文件夹是常规做法，并非特殊的注册表：托管的智能代理将普通项目文件组织在内，因此代理导入的任何本地模块都能正常工作。当这些文件夹存在时，命令行工具会将剩余文件视为托管的系统提示（instructions.md）、身份标识（identity.*）、连接器（connectors/**）、消息通道（channels/**）、定时任务（schedules/**）、技能（skills/**）、沙箱配置（sandbox/）以及本地 Harbor 评测任务（evals/）。 The only required file is agent.ts or agent.tsx at the project root. It must export a named agent created with defineDeepAgent. Use only one agent entry in a project. See Agent definition.

​
How MDA treats project files

  • Managed context: instructions.md defines the system prompt. Each directory under skills/ contains task-specific instructions. MDA syncs both to Context Hub.
  • Application code: Files under tools/ and middleware/ are ordinary project modules. Import them from the agent entry. Other local modules work the same way.
  • Managed configuration: Root identity.ts and memory.ts, direct children of channels/ and schedules/, and sandbox/index.ts enable their corresponding capabilities.
  • Dependencies and secrets: Declare dependencies in package.json. MDA loads .env locally and forwards eligible values as deployment secrets, but never includes .env files in the build archive.
  • Evals: Managed Deep Agents evals are Harbor evals. evals/tasks/ is the canonical Harbor task dataset. Author tasks there directly, or run mda evals init <name> to create an optional starter under evals/scaffold/. mda evals compile copies scaffolds into evals/tasks/ and packages the agent for Harbor. The evals/ directory is not included in the deployed agent build.
The layout above shows the common .ts names. TypeScript managed declarations also accept the supported .tsx, .mts, or .cts variants.
Connect these docs to Claude, VSCode, and more via MCP for real-time answers.
Edit this page on GitHub or file an issue.