GILD verdict
Choose LangGraph when agents branch, loop, wait for people, and must survive restarts, as GitLab's Duo Agent Platform and LinkedIn's Hiring Assistant do, because it saves state at every step and leaves every transition to your team.
For linear pipelines it adds overhead, which is why yeos.ai dropped it in July 2026. Pin versions, plan checkpoint retention and decide hosting before you build on it.
LangGraph
Agent Framework by LangChain
+25 PositiveFit and limits
Who it fits
- An agent has branches, retries, or loops, and the team wants its own code to set the path each run takes.
- Runs last minutes or days, wait for human approval, or must resume after a crash from the last completed step.
- The team wants one model-agnostic runtime for several agents, in Python or TypeScript, with streaming to a UI.
- The team already uses LangSmith, or wants managed hosting from the same vendor through LangSmith Deployment.
- A platform team wants one runtime under many agent flows: GitLab compiles YAML flow definitions into LangGraph graphs for Duo Agent Platform, which it says cut its own code per flow by 45% (Sep 22, 2026).
Where it falls short
- A linear pipeline or a single tool-calling loop gains little: one engineer replaced inherited LangGraph code with plain Python and cut over 1,000 lines without changing behavior.
- The learning curve is steep, and graph state is loosely typed by default, which shows up as runtime errors and hard-to-trace data flow in large graphs.
- Before 1.0, releases broke code and checkpointer packages regressed. 1.x promises stability, yet langchain-core 1.7.0 on October 22, 2026 changed how tool-call arguments stream, and LangChain told teams to upgrade clients first or pin below it. Teams on 0.x must upgrade before 0.4 fixes end in December 2026.
- The in-memory dev server is not for production, so a team either pays for LangSmith Deployment or runs Postgres, workers, and scaling itself. Builders who run their own workers report building leases and fencing so a slow worker cannot write a stale checkpoint (Reddit, Oct 27, 2026).
- The open-source Postgres checkpointer never expires a checkpoint. Tiendanube's copilot leaves about 93 records per conversation at more than 120,000 conversations a week, so the team built its own partitioned cleanup job and a weekly export to S3 (Apr 28, 2026).
Implementation reality
Design the state before the graph
Define the state schema as a TypedDict, dataclass, or Pydantic model, with reducers for fields that several nodes update, such as message lists. Keep the state small and serializable, because every checkpoint stores it. LangGraph's docs say adding or removing state keys is safe for saved threads. Renaming or removing a node breaks threads paused on an interrupt, and tightening a field breaks old checkpoints. Add fields as optional and keep removed ones for one drain cycle.
Choose a production checkpointer
Use the Postgres or MongoDB checkpointer in production; the in-memory saver loses everything on restart. Pick a durability mode per workload: sync writes every checkpoint before the next step, async overlaps writes with execution, and exit writes only when the run finishes.
Put people where actions are irreversible
Call interrupt() before sends, payments, deletions, and writes to other systems, and resume with Command(resume=...). Time travel replays or forks a run from a saved checkpoint, but every later node re-executes, including model calls and side effects, so make side-effecting nodes idempotent.
Decide where it runs
LangSmith Deployment runs graphs on Agent Server with checkpoints, threads, streaming, and autoscaling managed, from US$39 per seat per month on Plus plus metered compute. At LangChain's published rates, a Small Dedicated deployment that scales to its ceiling of 10 API server and 10 queue worker replicas costs about US$1,430 a month. The open-source library also runs inside your own FastAPI or Node service, where your team owns Postgres, workers, and cancellation handling.
What to measure in your pilot
- Run your test suite in CI against each new LangGraph and langchain-core minor release before upgrading production; more than one breaking minor a quarter means pinning exact versions.
- Record API server and queue worker replica counts on Dedicated through a week of peak traffic; if Small sits near its 10-replica ceiling, price Medium before launch.
- Before each state schema change, replay 50 saved threads, interrupted ones included, on a staging copy of the new graph; any thread that fails to load means shipping the change as add-then-remove.
Sources
- Official documentation
- 20
- Public source
- 1
- LangGraph docs: Overview
- LangGraph product page and FAQ (MIT license)
- LangGraph on GitHub (MIT license, stars, issues)
- LangGraph docs: Graph API
- LangGraph docs: Functional API
- LangGraph docs: Persistence
- LangGraph docs: Checkpointers and durability modes
- LangGraph docs: Interrupts
- LangGraph docs: Use time travel
- LangGraph docs: Streaming
- LangGraph docs: Run a local server
- LangGraph docs: Studio
- LangChain docs: Release policy (LangGraph LTS)
- LangChain blog: LangChain and LangGraph reach v1.0 (Oct 22, 2025)
- LangSmith plans and pricing (LangSmith Deployment rates)
- LangSmith docs: LangSmith Deployment
- LangSmith docs: Cloud platform features (deployment types and sizes)
- LangSmith docs: Manage billing (LangSmith Deployment billing)
- LangChain docs: Changelog (langchain-core 1.7.0 breaking change, Oct 22, 2026)
- LangGraph docs: Backward compatibility for saved checkpoints
- arXiv: Amini et al. (Oracle), Open Agent Specification (Agent Spec), SimpleQA Verified results (Nov 7, 2025)