AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Deep Agents Deploy: an open alternative to Claude Managed Agents

Collected Oct 13, 2026

LangChain has launched Deep Agents Deploy in beta, a tool that takes a custom agent built on its open source Deep Agents harness and stands up the production infrastructure for it with a single command, deepagents deploy. The company positions the release as an open alternative to recently launched managed agent offerings such as Claude Managed Agents.

The pitch rests on what the company calls harness engineering, the practice of building the orchestration layer that turns a large language model into a working agent. A harness carries orchestration logic plus tools and skills, and builders extend it with their own instructions, tools and skills. Going to production with such a harness has traditionally involved several separate chores: deploying orchestration logic and memory in a multi-tenant, scalable way; spinning up a sandbox per agent session; and standing up endpoints for agent interaction, including human-in-the-loop and memory access. Deep Agents Deploy bundles those steps into one command.

When a team runs that command, it deploys a customized agent defined by a handful of parameters. The model parameter selects the underlying LLM; LangChain says Deep Agents works with any model or provider, listing OpenAI, Google, Anthropic, Azure, Bedrock, Fireworks, Baseten, Open Router and Ollama. AGENTS.md holds the core instruction set loaded at the start of a session. Skills supply specialized knowledge through markdown files and actions through scripts. An mcp.json file declares tools the agent can call over the MCP protocol using HTTPS or SSE. A sandbox parameter is optional and specifies a working environment for running skills; out-of-the-box integrations cover Daytona, Runloop, Modal and LangSmith Sandboxes, and LangChain says any sandbox provider can be used.

Under the hood, the command bundles the agent with its own LangSmith Deployment server, described as production ready and horizontally scalable. That server exposes more than 30 endpoints. Among them are MCP, for calling a deployed agent as a tool; A2A, for multi-agent setups; Agent Protocol, for building UIs that interact with the agent; human-in-the-loop endpoints, for placing guardrails on what the agent may do without human intervention; and memory endpoints, for reading short-term or long-term memory. LangSmith Deployments can be self-hosted.

The open ecosystem angle is deliberate. The underlying deepagents harness is open source under the MIT license and available in both Python and TypeScript. Agent instructions use AGENTS.md, an open standard, and specialized knowledge uses Agent Skills, also an open standard. The project integrates with every model provider, letting teams pick a combination per task, including open models. Agents are exposed through MCP, A2A and Agent Protocol.

LangChain's central argument against proprietary managed agents concerns memory. Because a harness manages context, and memory is a form of context, the company argues that as more of a harness becomes closed and sits behind an API, so does the agent's memory. Switching models, in this view, is comparatively easy; bundling short- or long-term memory behind an API creates far more lock-in. The illustrative case is an agent that learns over time, such as an internal SDR agent or a customer-facing sales agent. Its accumulated memory is part of a data flywheel, but if that memory lives behind a closed API, leaving means resetting the agent's memory and starting over. Deep Agents Deploy stores memory in standard formats including AGENTS.md, skills and other files, allows direct querying through an API, and, when self-hosted, keeps memory in the customer's own databases.

My reading is that the comparison LangChain draws against Claude Managed Agents is architectural rather than feature-for-feature: the harness, agent server and sandbox layers are broadly similar in shape, and the differentiation is about who controls memory and where the escape hatches are. The trade-off is familiar from open infrastructure generally. Managed offerings typically reduce operational burden; self-hosting and owning memory means owning persistence, scaling, and availability, even if the deployment server itself is packaged. Teams should also note the beta label, which suggests APIs and defaults may still move.

Why it matters: developers and product teams building agents now have a deployment path that keeps the harness open source, model agnostic and, if self-hosted, keeps accumulated agent memory inside their own infrastructure. For teams whose agents learn from user interactions, that memory is often the most valuable asset, so the choice between a managed platform and an open harness is increasingly a data-ownership decision rather than a purely technical one. The beta status is the main caveat to weigh against production plans.

Read at LangChain

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Today we’re launching Deep Agents deploy in beta. Deep Agents deploy is the fastest way to deploy a model agnostic, open source agent harness in a production ready way.