THE DUDE’S FIELD GUIDE
LangGraph vs CrewAI: choose an agent workflow framework
The quick answer
Consider LangGraph when explicit state, durable execution and human review are central to the workflow. Consider CrewAI when role-based agents are a useful way to organize the work, with Flows controlling the surrounding process. Start with ordinary code if the steps are already predictable. Neither framework guarantees better answers or free model inference.
This comparison summarizes documented capabilities. We have not run a comparative benchmark; fit depends on your workload and environment.
At-a-glance comparison
| Project | Best fit | Costs and services | Why consider it | Watch out for |
|---|---|---|---|---|
| langchain-ai/langgraph | Stateful workflows with explicit orchestration | Model inference, persistence and hosting depend on your configuration | Durable execution and human-in-the-loop primitives | You still design state, recovery and side effects |
| crewAIInc/crewAI | Role-based teams within a controlled workflow | Model and tool providers may charge; hosted products are separate | Crews organize agent tasks; Flows manage execution paths | More agents can add calls and coordination without improving results |
Choose the unit of control
LangGraph describes itself as infrastructure for long-running, stateful agents and workflows. Its documented capabilities include persistence and human intervention. CrewAI offers complementary abstractions: Crews for collaboration among role-based agents and Flows for event-driven execution. Our practical recommendation is to sketch the workflow before choosing a library. If the sketch emphasizes state transitions and interruptions, investigate LangGraph first. If specialized roles help explain the work, investigate CrewAI, while keeping the outer process explicit. This is a fit assessment, not a performance ranking.
A concrete example: a repository research brief
Suppose you need to retrieve a README, check its license, draft a summary and ask a person to approve publication. Retrieval and license parsing can be ordinary functions; drafting may benefit from a model. Define the expected output for every step and reject missing evidence. A research agent and an editing agent are optional implementation choices, not requirements. Compare either framework against this simpler baseline before introducing several agents. Never treat agreement between agents as independent verification of a claim.
Plan recovery before deployment
CrewAI's production architecture guidance starts with a Flow, explicit state, focused Crews and structured task outputs. It also documents persisted Flow state. LangGraph advertises durable execution and human review as core capabilities. Those features do not, by themselves, establish that an external action will happen only once. For your evaluation, stop the process between drafting and publication, restart it, and inspect what repeats. Require a clear approval checkpoint and a way to recognize an action that already completed.
Budget for the whole workflow
Separate framework code from model inference, external tools, storage and hosting. A locally running Python process can still call a remote model or a paid search service. Read the selected provider configuration and deployment documentation before describing a workflow as local or free. For the research-brief example, record calls, retries and total cost per accepted brief. A cheaper individual model call is not necessarily a cheaper workflow if it needs more corrections. Do not assume that a hosted account or subscription is included with the repository.
Use the same evaluation tasks
Choose ten representative inputs, including an inaccessible repository, an ambiguous license, a changed README and an interruption requiring review. Decide what a correct result looks like before running either implementation. Measure supported claims, invalid outputs, duplicate actions, elapsed time, model cost and manual correction time. Keep the model, source material and permission boundaries comparable. Record failures as well as successful examples. Pick the smallest implementation that meets your requirements; our documentation review cannot predict which will win on your workload.
Common questions
Is CrewAI only for autonomous teams?
No. CrewAI also provides Flows for explicit, event-driven control, and its production guidance recommends using a Flow as the application's entry point. A Flow can delegate a focused task to a Crew.
Does LangGraph require a multi-agent design?
No. Its stated scope includes stateful workflows and agents. A single agent or a workflow combining ordinary functions with a model can be a sensible starting point.
Which framework is cheaper?
There is no workload-independent answer. Compare the complete run, including inference, tool calls, retries, persistence and human corrections. This guide reports documented capabilities, not measured cost or accuracy benchmarks.
Sources and review
- langchain-ai/langgraph documentation
- crewAIInc/crewAI documentation
- CrewAI production architecture, version 1.15.0
Check current installation instructions, licenses and service terms before choosing a tool. This guide’s review date changes only after an editorial recheck.
How we research and label evidence · Explore AI agents & orchestration