Skip to main content

Buyer guide

LangSmith Deployment: Choosing How Your Agents Run

Compare Serverless with Dedicated and Cloud with hybrid or standalone Agent Server. Price each option in LSUs before customers use the agent.

Reading time
3 min read
Sources
9 cited
In this piece
2 products

Run customer-facing agents on Dedicated and background agents on Serverless, and price the size in LSUs before launch. By GILD's calculation, a Small Serverless deployment left running costs about US$52 a month and a Small Dedicated deployment about US$384. Agent Server runs long-lived agents with threads, saved state, streaming and cron schedules, in LangChain's cloud, hybrid, self-hosted, or standalone.

When you need this

  • An agent built on LangGraph, or on another framework, is ready for users and needs durable execution and streaming.
  • Your organization was on the earlier per-run and uptime deployment pricing, which moved to usage-based billing on October 1, 2026.
  • Security requires agents to run inside your own network.

Evaluation criteria

  1. Framework

    Can the agent run on Agent Server?

    Look forLangSmith Deployment runs agents from any framework. LangGraph and LangChain deploy through the LangGraph CLI, Google ADK through the deployments-wrap-sdk package, and Claude Agent SDK, Strands, CrewAI, and AutoGen through the Functional API or the same package.

  2. Deployment type

    Is the agent in the critical path?

    Look forServerless runs on shared infrastructure and suits background and development agents that can tolerate latency. Its scale-to-zero is in beta. For customer-facing agents, LangChain recommends Dedicated, which stays on with its own database, automatic backups, and high availability.

  3. Size and cost

    What does a month cost?

    Look forRuntime compute bills 0.0675 LSU per vCPU-hour and memory 0.009 LSU per GiB-hour. A Dedicated database adds 0.177 LSU per vCPU-hour and 0.025 per GiB-hour, at US$1 per LSU. For a full 720-hour month, GILD calculates about US$52 for a Small Serverless deployment (0.8 vCPU, 2 GiB, 0.072 LSU an hour) left running. The same calculation gives US$384 for a Small Dedicated deployment (3 vCPU and 6 GiB runtime, 1 vCPU and 4 GiB database). Plus includes 1 free Small Serverless deployment.

  4. Where it runs

    Who runs the control plane and the Agent Servers?

    Look forOn Cloud, LangChain operates both on AWS and GCP from a connected GitHub repository, on Plus or above. Hybrid keeps Agent Servers in your infrastructure under LangChain's control plane. Enterprise self-hosting puts both in your Kubernetes cluster. Standalone runs Agent Server containers on Docker or Kubernetes with your PostgreSQL, Redis, and LangSmith license, without a control plane.

  5. Limits

    What will the platform reject?

    Look forCloud deployments reject request payloads over 25 MB. Dedicated deployments add replicas based on CPU, memory, and pending runs. When a replica stays under pressure, move up a size.

  6. Data in the runtime

    What does the Agent Server store?

    Look forAgent Server stores checkpoints, long-term memories, threads, and assistants in PostgreSQL. Set LANGGRAPH_AES_KEY to encrypt checkpoints, set TTLs in langgraph.json, and turn tracing off with LANGSMITH_TRACING=false when an agent's data must not reach LangSmith.

  7. Team experience

    What do teams report?

    Look forA developer who ran the product for a year, from its LangGraph Platform days, called it rough early, with weak monorepo support. The same developer said the benefits outweighed the costs once it matured.

Questions to ask vendors

  1. Is each agent customer-facing, and so Dedicated, or background work that can scale to zero?
  2. Which size does a load test show we need, and what is that per month in LSUs?
  3. Must the Agent Servers run in our network, and do we then need Enterprise?
  4. Which agents can send traces to LangSmith, and which must run with tracing off?

Red flags

  • A customer-facing agent on Serverless, where a cold start sits in front of the first request.
  • Dedicated deployments left running for development branches.
  • Budgets built on the pre-October 2026 per-run and uptime pricing.
  • Checkpoints with customer data stored unencrypted.

Products in this piece

Sources

Official documentation
8
Public source
1
  1. LangSmith docs: LangSmith DeploymentOfficial documentationdocs.langchain.com
  2. LangSmith docs: Cloud platform features (deployment types and sizes)Official documentationdocs.langchain.com
  3. LangSmith docs: Manage billing and spend limitsOfficial documentationdocs.langchain.com
  4. LangSmith plans and pricingOfficial documentationlangchain.com
  5. LangSmith docs: HybridOfficial documentationdocs.langchain.com
  6. LangSmith docs: Standalone Agent ServerOfficial documentationdocs.langchain.com
  7. LangSmith docs: Data storage and privacyOfficial documentationdocs.langchain.com
  8. LangSmith docs: StudioOfficial documentationdocs.langchain.com
  9. Hacker News comment on a year of LangGraph Platform, now LangSmith Deployment (Oct 21, 2025)Public sourcenews.ycombinator.com

Keep reading

  1. Buyer guideRunning LangGraph in Production: LangSmith Deployment or Your Own ServersPrice each LangSmith Deployment size in LSUs. Compare a standalone Agent Server with your own service and assign responsibility for operations.
  2. Buyer guideLangGraph Production PlaybookEight production checks for a LangGraph agent leaving the in-memory saver for real users, including how Airbnb pauses a run for days until a person decides.
  3. Buyer guideLangGraph Version Pinning & Upgrade ChecklistHow long LangChain supports LangGraph 1.x and 0.4, and the two October 2026 security advisories that set minimum package versions.
Read the research