LangChain launches Managed Deep Agents public beta

LangChain has released Managed Deep Agents into public beta, offering developers a way to transition AI agents from local prototypes to production-scale deployments. Built on top of the open-source Deep Agents harness, this new managed service offloads the complex infrastructure management required to run persistent, long-running agents. Developers can write their agents in Python or TypeScript, test them locally, and deploy them to a hosted LangSmith environment using a single command.

To get started, developers can install the command-line tool using the uv tool install managed-deepagents command for Python or npm install -g managed-deepagents for TypeScript. Running mda dev starts a local development environment in LangSmith Studio, while mda deploy compiles the project and uploads it to a hosted LangSmith deployment. The service manages the operational complexities that typically break traditional stateless infrastructure, such as durable execution for multi-day tasks, streaming progress updates, and thread state persistence across restarts.

The platform introduces several built-in primitives to enhance agent capabilities. It features native support for LangSmith Sandboxes, allowing agents to securely execute code and manipulate files in isolated environments. For testing, Managed Deep Agents integrates with Harbor, using commands like mda evals init and mda evals compile to package agents for state-based evaluation. Additionally, developers can connect their agents directly to communication platforms like Slack or GitHub by adding configuration files to a dedicated channels directory.

For long-term interactions, the runtime provides agent-scoped memory backed by LangSmith's Context Hub, preserving learned information across redeployments. It also includes an identity system that integrates with OIDC providers to scope threads per user ID. For AI practitioners, this release eliminates the months of engineering time typically spent building custom infrastructure for agent persistence, sandboxing, and human-in-the-loop approvals, allowing teams to focus entirely on defining agent behavior and domain logic.