LangChain Managed Deep Agents
LangSmith managed runtime plus Connections for agent-owned vs caller-owned credentials
Managed Deep Agents (MDA) lets you define an agent as a project folder and run it on LangSmith’s managed Agent Server: durable sessions, sandboxes, Context Hub, channels, schedules, and observability. Connections are named credentials stored in the LangSmith workspace, retrieved at runtime with connections.get(). Ownership is agent-owned (shared by every caller) or caller/user-owned (resolved per person; may pause for OAuth). Ownership is fixed at create time. Public docs describe MDA as a public beta on LangSmith Cloud. Pair Connections with the Agent Payments playbook when the credential is a spend capability.
Use Cases
- Deploy a deep agent folder to LangSmith without running your own agent server
- Store API keys as agent-owned Connections instead of baking them into the build
- Resolve per-caller OAuth so actions carry the caller’s identity
- Pause a run for authorization when the caller has no grant yet
- Keep credential rotation in the workspace, not in a redeploy
Key Features
Managed Agent Server
LangSmith runs the Deep Agents harness: durable threads, checkpointing, streaming, sandboxes, channels, and traces.
Project-as-folder
Business logic lives in the repo (instructions, tools, skills, connectors, middleware). The managed runtime hosts it.
Connections ownership
Agent-owned credentials are shared across callers. Caller/user-owned credentials resolve per person. Set ownership at create; it does not change later.
OAuth handled by the runtime
Managed Deep Agents runs the authorization round-trip. Docs describe no callback route or token store in your project for catalog providers.
MCP connectors
Remote MCP tools via connectors/ (HTTP or SSE). Stdio MCP is not supported on the managed path per LangChain docs.