Managed Deep Agents is now in public beta. Go from prototype to production scale

LangChain's Managed Deep Agents, in public beta since August 7, hosts agents built on its open-source Deep Agents harness so a team does not have to build the servers that keep long-running agents alive. A harness is the code around a model that runs an agent: calling tools, keeping working files, managing a growing context, delegating to subagents, loading skills and pausing for human approval. Deep Agents packages that pattern and is model agnostic, so the team chooses the model. According to the launch post, a developer writes the agent in Python or TypeScript as a project folder holding its instructions, tools, skills, memory, identity rules, channels and evals, tests it locally with mda dev, and runs mda deploy to compile it and create a hosted deployment on LangSmith, the company's agent platform.

The hosting matters because agents break the assumptions of ordinary web infrastructure, which expects short requests that keep no state. LangChain notes that an agent may run for minutes, hours or days, wait for approval, resume after a user replies and need to survive a server restart. The managed runtime, built on the Agent Server already used by LangSmith Deployment, supplies durable execution so a run can pause, retry and resume without losing work, and keeps conversation state across turns and failures. By default each conversation thread gets its own isolated sandbox for files and code. Memory files survive redeploys, so updating the agent's instructions does not wipe what it has stored. A Slack channel file mounts the event endpoint and verifies the provider's signatures, and an OIDC identity provider lets the runtime keep each end user's threads separate. For testing, it packages the agent for Harbor, a framework that runs a task in an isolated environment and grades the resulting files or state, though teams still run Harbor themselves.

The tradeoff is an opinionated runtime in exchange for less infrastructure work. LangChain says teams that need custom routes, application code beside the agent, custom auth logic or direct control of the persistence layer should use LangSmith Deployment directly, or run the open-source harness on their own infrastructure. The beta also has limits: it runs only in LangSmith Cloud's US region, it is driven from the command line while the API is still being finalized, and the identity model is described as basic: for now, an agent can run with a fixed set of credentials. Model choice stays open, but during the beta the runtime, thread state and memory live in LangSmith's US cloud. The customer quotes in the post are testimonials rather than measured results.