Managed Deep Agents

Managed Deep Agents lets you define an agent as a folder and run it on managed LangSmith infrastructure. You provide the business logic, and Managed Deep Agents provides the agent harness and production infrastructure.

​

Define your agent

An agent starts as a project folder that contains the business logic for its behavior:

• Instructions: The prompt that defines what the agent does and how it behaves.
• Tools: Functions the agent can call to interact with other systems or take actions.
• Skills: Reusable, task-specific instructions and resources.

You can add other capabilities as needed. For the complete folder layout, see Project structure.

​

Run on a managed harness

Managed Deep Agents combines three layers:

• Your business logic: The instructions, tools, and skills in your project folder.
• Agent harness: The battle-tested Deep Agents harness that runs the agent and connects its business logic.
• Managed infrastructure: LangSmith infrastructure that operates the agent at scale for production and multi-user applications.

This separation lets you focus on what the agent should do instead of building and operating the systems required to run it.

​

Managed infrastructure

The opinionated infrastructure consists of several pieces:

• Runtime: LangSmith Agent Server runs agents in a durable, fault-tolerant manner.
• Sandboxes: LangSmith Sandboxes let agents write and execute untrusted code in an isolated environment.
• Evals: Managed Deep Agents uses Harbor tasks to test agent behavior.
• Channels: The channels abstraction connects an agent to platforms where its users work.
• Memory: Managed memory lets agents remember information across interactions.
• Context management: LangSmith Context Hub manages agent instructions and skills. You can update them in the LangSmith UI without redeploying the agent.

To create and deploy an agent, follow the Managed Deep Agents quickstart.

---