LangChain launches Deep Agents deploy, an open rival to Claude Managed Agents

One command for the production steps

According to LangChain, "harness engineering" has become the discipline of building harnesses that turn LLMs into agents. These harnesses hold orchestration logic, tools and skills, and builders can add custom instructions, tools and skills for their own use case.

Going to production, the company says, requires several steps:

• Deploy the agent orchestration logic and memory in a multi-tenant, scalable way.
• Set up sandboxes so they are spun up per agent session.
• Stand up endpoints to interact with the agent, covering MCP, A2A and others for human-in-the-loop, memory and more.

Deep Agents deploy bundles all of these into one command: deepagents deploy. Under the hood, it packages the user's Deep Agent with its own LangSmith Deployment server, which LangChain describes as production ready and horizontally scalable.

What gets deployed and what the server exposes

Users specify a few parameters for their custom agent:

• model: the large language model to use. Deep Agents works with any model or provider, including OpenAI, Google, Anthropic, Azure, Bedrock, Fireworks, Baseten, Open Router and Ollama.
• AGENTS.md: the core instruction set, loaded at the start of the session.
• skills: Agent Skills that provide specialized knowledge through markdown files and actions through scripts.
• mcp.json: tools the agent can call over the MCP protocol (HTTPS/SSE).
• sandbox: an optional environment where the agent can work and run skills. Integrations with Daytona, Runloop, Modal and LangSmith Sandboxes are included, and any sandbox provider can be used.

The deployed server has 30+ endpoints. They include MCP, so deployed agents can be called as tools, and A2A, for multi-agent setups. Others cover Agent Protocol, for building interfaces to the agent, human-in-the-loop guardrails, and memory endpoints for short term or long term memory.

The open ecosystem argument against lock-in

LangChain builds the product on deepagents, an MIT-licensed open source harness available for Python and TypeScript. It uses the open standards AGENTS.md and Agent Skills, and exposes agents through MCP, A2A and Agent Protocol. The company says it integrates with every model provider and every sandbox provider, and states: "No Anthropic lock-in." LangSmith Deployments can also be self-hosted.

On the comparison with Claude Managed Agents, LangChain says the high-level architecture of harness, agent server and sandboxes is the same. It calls Claude Managed Agents a "walled garden" that creates lock-in.

The central point is memory. LangChain argues that a harness manages context, and memory is context, so closing parts of the harness behind an API also closes the memory. The company says switching between models is fairly easy, but bundling memory behind model APIs creates heavy lock-in. It gives the example of an internal SDR agent that learns from users: moving off the harness or model would mean resetting that memory and starting from scratch. For customer-facing agents, it says, the accumulated memories would belong to whoever owns the closed API.

What it means for teams building agents

LangChain says deepagents deploy stores memory in a standard format (AGENTS.md, skills and other files) and lets users query it directly via API. If a team self-hosts, the company says, memory stays in its own databases only.

For teams weighing hosted agent platforms, the pitch is a choice of models and sandboxes while keeping ownership of what their agents learn. The product is in beta, and LangChain points to its documentation and GitHub for those who want to try it.