LangChain released Managed Deep Agents in public beta, a hosted runtime for long-running agentic workloads built on its open-source Deep Agents harness. The service handles production infrastructure (durable execution, persistence, sandboxes, streaming) while keeping agent logic (prompts, tools, middleware, identity, evals) under user control. Developers author agents in Python or TypeScript, test locally via LangSmith Studio, and deploy with a single `mda deploy` command.
The key operational primitives Managed Deep Agents provides: durable execution so long-running agents pause/resume without losing state; sandboxes for file work and code execution (with first-class LangSmith sandbox support); persistence so thread state survives restarts; streaming so users see progress in real time; and evals so teams can test behavior and tool use before deployment. The runtime also includes identity/auth, memory scoping, channel integrations (Slack, GitHub), and managed cron schedules.
Deep Agents is model-agnostic and open-source, allowing users to bring their own models, instructions, and business logic. The project structure is code-first—agent.py, instructions.md, tools/, channels/, middleware/, skills/, sandbox/, and evals/ directories—making it easy to organize all agent primitives. Managed Deep Agents compiles projects, syncs context to LangSmith's Context Hub, and creates a hosted deployment.
For teams building production agents, this announcement addresses the operational gap between prototype and scale. Agent infrastructure has been a months-to-quarters engineering lift: durable threads, persistent memory, cancellation, retry logic, human approval gates, and failure recovery are now table stakes for product agents. LangChain's move to absorb this burden behind a managed service frees engineering to focus on agent behavior, prompt optimization, and tool design. Watch for competitive moves from Anthropic Workbench and OpenAI Agent platforms to bundle similar infrastructure.