LangChain Opens Managed Deep Agents Public Beta
Python and TypeScript teams can move a code-first agent into a hosted LangSmith runtime that supplies durable execution, sandboxes, memory, identity and evaluation plumbing.
LangChain opened Managed Deep Agents in public beta on August 7, giving Python and TypeScript teams a managed path from a locally tested Deep Agent to a hosted LangSmith runtime. Developers retain control of the model, instructions, tools, middleware and subagents, while LangSmith handles persistence, memory mounts, skill loading, sandbox lifecycle and deployment, according to the official announcement.
The release matters because long-running agents need more than a model call. They must keep state, recover from interruptions, isolate code execution and preserve user boundaries. Managed Deep Agents turns those recurring operational requirements into a product layer around LangChain's open-source, model-agnostic Deep Agents harness.
What changed in the public beta
A Managed Deep Agent is a code-first project stored in the developer's repository. LangChain provides command-line flows to create a project, run it locally in LangSmith Studio and deploy it to the managed runtime. The project can organize instructions, identity rules, memory behavior, custom tools, channels, middleware, schedules, connectors, skills, sandbox configuration and evaluations.
On deployment, the service compiles the project, synchronizes deployment-owned context to LangSmith Context Hub, uploads the build and creates a hosted LangSmith deployment. That separates the agent's application logic from the runtime work needed to keep executions available and inspectable.
What the runtime manages
The managed layer provides durable execution so a long task can pause, retry and resume without losing work. It also supplies streaming, persistent thread state, isolated sandboxes, evaluations, channels, memory and identity controls. LangChain says each durable thread receives its own sandbox by default, while developers can choose an agent-level scope when several threads should share one workspace.
That design is useful for agents that inspect files, run tests, invoke command-line tools or generate artifacts. Sandbox activity is traced in LangSmith, giving operators a record to inspect when a run succeeds or fails. The beta also supports channels such as Slack, allowing an agent to receive events and reply without a separately built integration service.
Evaluation, memory and user boundaries
Managed Deep Agents packages agents for Harbor-based evaluations, where tasks run in an isolated environment and verifiers can inspect resulting files or state. Harbor itself still runs separately, either in local Docker or another configured environment, so the beta provides the handoff rather than replacing the evaluation system.
Each deployment starts with agent-scoped memory backed by Context Hub. Developers define memory behavior in code, and redeployment preserves runtime-created memories while refreshing deployment-owned instructions and skills. For multi-user deployments, an optional OIDC provider can scope threads by end-user identity.
What remains unproven
This is a first-party public-beta announcement, not independent evidence of reliability, security or production performance. The fetched material does not establish pricing, regional availability, service-level commitments or comparative operating cost. Teams evaluating the beta should test recovery behavior, sandbox isolation, identity boundaries and upgrade workflows against their own requirements before treating the managed runtime as production-ready.
Status
Confirmed product beta. Internal confidence is medium because availability and features come from LangChain's official announcement without independent operational testing in the fetched evidence.
Sources
Update note: Last reviewed 2026-08-13. We will revise this post if LangChain changes beta access, runtime capabilities or supporting evidence.
Sources
- LangChain — Managed Deep Agents public beta — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.