Quick answer
Choose Managed Deep Agents for the shortest opinionated path from a Deep Agents project to a LangSmith-hosted runtime. Choose LangSmith Deployment directly when the application needs custom routes, code alongside the graph, custom auth logic, or direct persistence control. Choose the open-source Deep Agents harness on your own infrastructure when runtime placement and operations must remain under your control.
This is an ownership decision, not a feature checklist. The safest choice is the one whose control boundary matches the team’s compliance, platform, incident-response, and engineering capacity.
Decision matrix
| Question | Managed Deep Agents | LangSmith Deployment directly | Self-hosted Deep Agents harness |
|---|---|---|---|
| Harness | Deep Agents supplied as part of the managed path | You compose the agent application | You own and run the open-source harness |
| Runtime owner | LangSmith | LangSmith deployment infrastructure with more application control | Your team or infrastructure provider |
| Configuration | Opinionated project folder and mda CLI | Lower-level deployment configuration | Your code, packaging, orchestration, and operations |
| Best fit | Fast hosted Deep Agent delivery | Custom app and graph deployment needs | Runtime, network, data, or platform control |
| Current availability | Public beta, LangSmith Cloud US only, CLI-first | Verify current LangSmith plan and deployment options | Depends on your chosen infrastructure and dependencies |
| Persistence and recovery | Managed production primitives | Configured through the deployment architecture | Designed, operated, tested, and repaired by your team |
| Auth flexibility | Current documented defaults and supported identity path | Better fit for custom auth logic | Fully your responsibility |
| Operational burden | Lowest of the three, not zero | Medium and architecture-dependent | Highest |
Choose Managed Deep Agents when defaults fit
The managed path is compelling when the product is genuinely a Deep Agent, the team wants code-first ownership of behavior, and LangSmith can own the runtime scaffolding. Durable execution, persistence, streaming, sandboxes, memory, channels, schedules, traces, and eval packaging can shorten delivery.
The tradeoff is an opinionated beta surface. Confirm US-region acceptability, supported language and package versions, CLI automation, identity design, memory scope, sandbox policy, price, quotas, support, and change management. Do not assume a future region, deployment method, API, or self-hosted option is available today.
Choose LangSmith Deployment when the application is broader
LangChain’s public-beta announcement explicitly points to LangSmith Deployment for custom routes, application code alongside the graph, custom authentication logic, or direct control over persistence. That is the clearer path when the agent is one part of a larger application server or when the opinionated Managed Deep Agents project model hides controls the application must own.
More flexibility also means more design work. Map deployment topology, durable state, authentication, resource authorization, migrations, data lifecycle, observability, rollback, and incident ownership before estimating effort.
Choose self-hosting when infrastructure control is the requirement
The Deep Agents harness is open source and model agnostic. Running it yourself can satisfy infrastructure placement, network, customization, or platform-standardization needs that the current managed beta cannot. It also transfers runtime availability, durable execution, persistence, queueing, sandbox provisioning, secrets, upgrades, scaling, tracing, deletion, and recovery to your team.
Self-hosting is not automatic independence. Inventory model APIs, databases, object storage, identity providers, tool systems, telemetry, and package supply chains. The runtime can be yours while important data and control paths remain external.
Run a reversible selection
Build the same narrow task and failure suite on the serious candidates. Compare successful completion, pause and resume, duplicate prevention, user isolation, memory behavior, sandbox cleanup, trace usefulness, deployment time, rollback time, and operator burden. Keep instructions, tools, eval fixtures, and business logic portable where practical.
For current setup boundaries, read the Managed Deep Agents public beta guide. For state and access controls, use the memory, sandboxes, and identity guide and the readiness checklist.
Frequently asked questions
When should a team choose Managed Deep Agents?
Choose it when the team wants the open-source Deep Agents harness with an opinionated LangSmith-managed runtime and can accept the current public-beta, US-region, and CLI-first boundaries.
When is direct LangSmith Deployment a better fit?
LangChain recommends direct LangSmith Deployment when you need custom routes, application code alongside the graph, custom authentication logic, or direct control over persistence.
Does self-hosting remove model-provider dependencies?
Not necessarily. Self-hosting the harness changes runtime ownership; model, tool, storage, identity, and other external dependencies still follow the architecture you choose.
Official sources
- LangChain: Managed Deep Agents public beta
- LangChain: Managed Deep Agents overview
- LangChain: Deep Agents overview
Source check: August 12, 2026. Recheck release stage, regions, deployment options, product limits, auth support, persistence controls, pricing, and support before choosing a platform.