Deep Agents Deploy: An Open Alternative to Claude Managed Agents

Deep Agents Deploy has been released in beta. Built on top of the model-agnostic and open source Deep Agents harness, its core value is bundling everything needed for production deployment into a single deepagents deploy command.

Building a production-ready agent usually requires preparing several separate steps. The author summarizes these as follows.

• Deploying the agent's orchestration logic and memory in a multitenant, scalable way
• Configuring a sandbox that spins up separately per session
• Providing endpoints for MCP, A2A, human-in-the-loop, and memory access

What gets deployed is a custom agent defined by the user. The items you can specify are model, AGENTS.md, skills, mcp.json, and a sandbox if needed, and the models are explained to support a variety of providers including OpenAI, Google, Anthropic, Azure, Bedrock, Fireworks, Baseten, Open Router, and Ollama.

Under the hood, it's combined with a LangSmith Deployment server, operating as a production-grade, horizontally scalable server that provides 30+ endpoints. These include endpoints related to MCP, A2A, Agent Protocol, human-in-the-loop, and memory.

The core message of this announcement is the open ecosystem. The harness uses the MIT-licensed deepagents, instructions are defined via AGENTS.md, and specialized functionality is defined via Agent Skills. Models and sandboxes are not tied to a specific vendor, and if you self-host LangSmith Deployments, memory can also be kept within your own database.

By contrast, the author argues that Claude Managed Agents, even with the same harness/server/sandbox structure, results in greater lock-in due to its closed environment. In particular, it points out that when memory is tied behind an API, this creates a switching cost even greater than changing models, and that for outward-facing agents, the memory and data flywheel that accumulates over time effectively becomes owned by the platform.