Managed Deep Agents Launches in Public Beta

Managed Deep Agents has launched in public beta. Developers can write Deep Agents in Python or TypeScript, test them locally, and deploy them to LangSmith's managed runtime with a single mda deploy command.

Developers directly control the core behavior of the agent, including models, instructions, tools, middleware, and sub-agents. LangSmith handles operational infrastructure such as persistence, memory mounts, skill loading, sandbox lifecycle, and deployment.

Deep Agents is an open-source, model-agnostic agent harness that provides common patterns such as tool calling, task file management, context management during long-running executions, sub-agent delegation, domain-specific skill loading, and human approval before sensitive operations.

Managed Deep Agents provides the features necessary to operate this harness in production environments.

• Durable execution: Supports pausing, retrying, and resuming long-running tasks
• Streaming: Transmits task progress in real time
• Persistence: Maintains thread state across conversation switches, restarts, and failures
• Sandboxes: Performs file operations, code execution, and CLI usage in isolated environments
• Evals: Evaluates agent behavior, tool usage, and state changes before and after deployment
• Channels: Connects with external tools such as Slack
• Memory: Stores persistent context and user preferences across conversations
• Identity: Manages user context and access scopes

The project manages agent code, dependencies, instructions, authentication and memory settings, tools, channels, middleware, schedules, connectors, skills, sandbox configurations, and evaluation code within a single directory structure. Running mda deploy builds the project, synchronizes the deployment target context to LangSmith Context Hub, and creates a hosted LangSmith deployment.