What Is Hosted Is the Runtime, Not the Framework
The Deep Agents framework itself stays open source and can still be run on your own infrastructure; what this beta hosts is the layer beneath it. The split of responsibility is explicit: the model, instructions, tools, middleware, subagents, and domain logic belong to the developer, while LangSmith owns persistence, memory mounts, skill loading, sandbox lifecycle, durable execution, streaming, authentication, scheduling, and tracing. The workflow runs through an `mda` CLI — `mda init` to scaffold a project, `mda dev` to run locally against LangSmith Studio, and `mda deploy` to ship, with evaluations handled by `mda evals init` and `compile`.
Sandboxes Isolated per Thread, in One Line of Configuration
Sandboxes are declared in code with `define_sandbox(provider="langsmith", scope="thread")`. By default each durable thread gets its own sandbox, which suits coding agents that need one workspace per user conversation or task; setting `scope="agent"` shares a single sandbox across threads instead. Sandbox activity flows back into LangSmith traces, so successful and failed runs can be inspected after the fact. None of this is trivial to build in-house — provisioning, cleanup, and crash recovery are each their own piece of engineering work.
Worth Evaluating, Not Yet Worth Migrating
The beta comes with real constraints. It runs only on LangSmith Cloud in the US region, with other regions and deployment methods promised for later. It is CLI-first while the APIs are still being finalized, the beta primitives may change based on feedback, and no pricing has been published. For teams outside the US, the region limit alone makes this a release to evaluate rather than migrate to: it is more useful right now to decide which capabilities are worth outsourcing to a managed runtime and which must stay in your own hands. Managed Deep Agents first appeared as a private beta at LangChain's Interrupt conference in May; this is the first time it has been open to everyone.
via: LangChain announcement: Managed Deep Agents is now in public beta, LangChain docs: Managed Deep Agents examples; verified 2026-08-09