November 13, 2026 · 6 min read
CrewAI and LangGraph are two AI agent frameworks. When an agent built with either settles into a routine, the fixed part can become a Whenever workflow: code steps for what repeats, one model call where judgment is needed, and triggers, secrets and run history on Whenever.
AI steps have no same-output guarantee and incur runtime AI usage when they run. Code-only steps keep following the rules you approve.
What CrewAI is, in its own words #
CrewAI’s documentation describes it as “the leading open-source framework for orchestrating autonomous AI agents and building complex workflows.” It has two building blocks. A crew is “a collaborative group of agents working together to achieve a set of tasks”, and Flows “allow you to create structured, event-driven workflows that manage state and control execution.”
A crew runs its tasks by a process. In the sequential process “tasks are executed one after another, allowing for a linear flow of work”; in the hierarchical process “a manager agent coordinates the crew, delegating tasks and validating outcomes before proceeding.” Calling kickoff() “starts the execution process according to the defined process flow.”
For hosting, CrewAI AMP (Agent Management Platform) “is where teams build, deploy, operate, and govern their Crews and Flows in production.” Its docs say every deployed Crew or Flow “becomes an automation with its own API, triggers, traces, and access rules”, started “on a schedule, from an app event, or through the REST API.”
What LangGraph is, and how it differs from LangChain #
LangChain’s documentation calls LangGraph “a low-level orchestration framework and runtime for long-running, stateful agents.” You “define workflows as graphs of shared state, nodes, and edges, and can mix deterministic steps with LLM-driven steps in the same graph.” The short version in the same docs: “nodes do the work; edges decide what runs next.”
A graph runs in super-steps: “Nodes that run in parallel share one super-step. Nodes that run in sequence use separate super-steps,” and “the run ends when every node is inactive and no messages remain in transit.” State outlives a run through checkpointers, which “persist a thread’s graph state as checkpoints”, and stores, which “persist application-defined data outside the graph state.”
LangGraph is not LangChain. The docs describe LangChain’s create_agent as “a minimal, highly configurable harness” around the model loop and note that “LangChain’s agents are built on top of LangGraph.” Their advice: “Use LangGraph, our low-level orchestration framework, for advanced needs combining deterministic and agentic workflows.” LangGraph “does not require LangChain.”
For hosting, LangSmith Agent Server is “a workflow orchestration runtime purpose-built for agent workloads”, offered as Cloud, Hybrid, self-hosted and standalone topologies. Its docs state that “Deployments on LangSmith Cloud require a Plus or Enterprise plan” and are “billed on the compute and database resources that each deployment requests.”
Which part of a framework agent repeats #
Read an agent that has been running for a month and the shape is usually the same. The trigger is fixed: a cron entry, or a request that arrives every morning. The fetch is fixed: the same feed, the same API, the same query. The post is fixed: the same Slack channel, the same inbox. The only part that needs a model is the middle, reading what came back and writing the result.
Both frameworks draw a similar line. CrewAI’s guidance is to “start with a Flow” and use a crew, within a Flow step, “when you need a team of agents to perform a specific, complex task that requires autonomy”; LangGraph lists “combine hand-coded nodes with model-driven nodes in one graph” among its features. The repeating part is the hand-coded part.
On Whenever that split is the whole design. The fixed steps become code that runs the same way every time, the judgment call stays a single model call inside one step, and the schedule, the secrets and the run history are the platform’s.
How a coding agent does the refactor #
Paste one line into Claude Code, Codex or Cursor:
Turn this task into a Whenever workflow and deploy it. Follow www.whenever.dev/skill.md
The skill tells the agent to read the source first and list its trigger, inputs, every API or tool it calls, and its outputs. Then it splits the fixed procedure from judgment: everything done the same way every run becomes a step that calls an integration, and only genuine judgment (classify, summarise, draft) becomes a model call inside a step. Pasting the whole prompt into one AI step is the agent again, not a workflow.
It maps when the job runs: a cron or interval becomes a schedule, and “when X happens” becomes a webhook or event. Credentials move out of the code. You connect each account once in Setup, and your coding agent never sees a secret.
Then it saves a draft through the Whenever MCP server. You read the draft’s graph (the trigger, each step, and where a model is called), run it once by hand, and publish to switch the trigger on. Every run is logged, and you get an email when an automated run fails.
A worked example #
The scheduled research brief workflow in the docs is the pattern end to end. A weekly trigger fires on Monday morning. A fetch step reads a changelog feed the owner sets in Setup, validates the shape of the reply and keeps at most twenty entries from the last seven days. One model call writes the brief from those entries, passed as data. A post step sends it to Slack and returns what Slack confirmed. A quiet week returns early without a model call, and a manual trigger beside the schedule lets you run the same code before publishing.
That is what most scheduled agents collapse into once the job is known: three steps, one of them a model call.
Deploy it, or keep it in the framework? #
Deploy it to Whenever when
- The job repeats and the steps are already known
- It reads and writes business tools through APIs or webhooks
- You want a run history someone else can read, and a model call only where you put one
Keep it in the framework when
- The path to the answer is open-ended, which is CrewAI’s own test for using a crew
- You need checkpoints, time travel or human-in-the-loop s inside a run, which LangGraph’s persistence is built for
- Every run needs a fresh plan rather than a fixed route
A few useful answers #
Is LangGraph the same as LangChain?
No. LangChain’s docs describe LangGraph as a low-level orchestration framework and LangChain’s create_agent as a harness built on top of it, and LangGraph “does not require LangChain.”
Do I need a framework at all?
Not for a job whose steps are known. A Whenever workflow is a module with a manifest of triggers, one step per unit of observable work and a run function that orchestrates them. The orchestration is code, so the only model call in a run is one the workflow makes on purpose.
Can the model still be called?
Yes. ai.generateText takes a prompt or a message history, and with tools it is one turn of an agent loop. The platform keeps no conversation and never runs a tool, so the loop, its turn cap and the argument checks are in your code.
What about state and memory?
A workflow run starts from its trigger and ends with what the run function returns; nothing replays a workflow and no step result is cached. Keep durable state in the systems the workflow already talks to, such as a sheet, a database read through the postgres port, or the thread it posts to. LangGraph’s checkpointers and stores, and a crew’s memory setting, are the frameworks’ answers to the same need; a workflow does not carry them over.
Sources & further reading #
Sources reviewed November 13, 2026.
- Introduction - CrewAI
- Crews - CrewAI
- Flows - CrewAI
- CrewAI AMP - CrewAI Platform
- LangGraph overview - Docs by LangChain
- Graph API overview - Docs by LangChain
- Persistence - Docs by LangChain
- LangChain overview - Docs by LangChain
- LangSmith Agent Server - Docs by LangChain
Choose your next step #
- Deploy from your coding agent
- Deploy a Claude agent or task as a workflow
- AI agent hosting without a VPS or a server to run
- A scheduled research brief workflow
Your first workflow is one sentence away. #
No credit card required. One workflow is free. AI for building has weekly usage limits.