LangGraph

LangGraph is for anyone burned by an agent that lost its place halfway through a long job. You draw the graph, decide which nodes are code and which are model calls, and the runtime keeps the state, so a run that stops for a human answer on Tuesday resumes on Wednesday where it left off. That is the daily trait.

It is low-level on purpose, and you will write more plumbing than you expected before the first agent does anything. Pick it for anything that runs longer than a request. Pick CrewAI for roles, less plumbing, and a crew by tonight.

The risk is a soft dependency. The graph runtime stands alone. The moment an agent needs a local model it imports ChatOllama from LangChain, and the moment it needs MCP tools it installs langchain[mcp], so the dependency you were told you did not need is the one that ships the tools. Version drift between the two lands on you.

The consequence: pin both packages together and treat a LangChain release as a LangGraph release. First-party MCP and Ollama adapters inside langgraph itself would change this verdict. What it does right: human-in-the-loop lets you inspect and change state at a checkpoint before a bad step lands.

1. A graph whose nodes may be hand-written code or model calls, so the deterministic parts stay deterministic and the model is confined to the nodes that need it. 2. Durable execution: state is persisted at each step and a run resumes from the last checkpoint after a crash, which makes a long run reproducible from its history. 3. Streaming. 4. Memory in working and cross-session tiers.

The consequence is that verification can be placed at any node as ordinary code, which is the property most agent frameworks lack. The documentation calls itself very low-level, which is accurate and rare.

Give away the orchestration, sell the observability and hosting. The Developer tier at 5k traces a month is a lead magnet, and the overview page lists Klarna, Uber and J.P. Morgan as users, the logo wall a Series C buys and a Series D sells. The library creates the traces; LangSmith charges for reading them.

Moat: the library's install base, which is switching cost only if teams adopt the hosted layer, and many will not. Likely acquirer: an observability incumbent, Datadog first, or a cloud that wants the funnel. Position: long, with the caveat that the library will outlive whatever happens to the company.

The demo is a workflow that pauses for approval and resumes. As a dependency: maintained by LangChain Inc, the library costs sixty engineers nothing, and it runs in whatever CI already runs Python. The paid path is LangSmith Deployment at $39 a seat on Plus, $2,340 a month for sixty, and SSO waits in the Enterprise tier at custom pricing.

Onboarding is the real cost: graphs are a new mental model and a mid-level engineer needs a week with the docs. Approved with conditions: the library now, LangSmith only after a retention review, because traces contain prompts and prompts contain data.

MIT and all Python I can read. The persistence layer is an interface, so my state lives in my own Postgres, not theirs, and a checkpointer is a class I can write in an afternoon. pip install "langchain[mcp]" turns any MCP server into graph tools, and the model layer takes whatever LangChain speaks, including my local box through ChatOllama.

The hosted platform is optional, which is the test I apply to every framework with a company behind it. If the company vanished, forty thousand stars' worth of people would fork it before lunch. This is mine to the extent code ever is.