CrewAI and FlowRunner

TL;DR

• FlowRunner adds the business side around a CrewAI crew: triggers from 2,100+ integrations, schedules, a review step that reaches a person by email, Slack, WhatsApp, phone, or a Form and waits for the answer, retries, and a record of every run your whole team can open.
• CrewAI is where you define agents, tasks, crews, and flows in Python. Keep it. FlowRunner doesn’t run Python, so the crew stays where it runs and the two talk over HTTP.
• CrewAI already ships human input (human_input, @human_feedback, and AMP’s webhook loop). FlowRunner plugs into those hooks as the receiving end, the part CrewAI leaves to you.
• You can build the FlowRunner side from a prompt with FlowRunner Prompt to Flow, in Claude Code or any MCP-capable AI tool.

What CrewAI is for

CrewAI is a Python framework for building systems of AI agents. Its GitHub description puts it as a “framework for orchestrating role-playing, autonomous AI agents.” It’s standalone, with its own primitives for agents, tasks, crews, flows, and tools, and it’s MIT licensed.

It has two building blocks. Flows are the structure of the application: steps, state, and branching. Crews are teams of agents, each with a role, a goal, and tools, that work through tasks together. CrewAI’s docs suggest starting with a Flow and adding a Crew wherever agent teamwork helps. Agents can use tools from MCP servers, and a Flow can save its state with @persist (SQLite by default) and pick up again under the same id.

For people in the loop, CrewAI has three options:

• human_input on a task. A person reviews the agent’s final answer before the task completes.
• @human_feedback in a Flow. By default it waits for console input. With emit, a model maps free-form feedback to named outcomes that route the flow. For anything asynchronous, you subclass HumanFeedbackProvider, notify an outside system, and raise HumanFeedbackPending. CrewAI saves the state, and later Flow.from_pending(flow_id) and resume() continue the run.
• The webhook loop on CrewAI AMP. Pass humanInputWebhook at kickoff, the execution pauses for human input, and you send the decision to AMP’s resume endpoint.

CrewAI AMP, the hosted platform, deploys crews from GitHub, Crew Studio, or the CLI, serves them through a REST API, and shows traces. It has app triggers for Gmail, Google Calendar, Google Drive, Outlook, OneDrive, Microsoft Teams, HubSpot, Salesforce, Slack, and Zapier. Its pricing page lists two plans: Basic, free with 50 workflow executions a month, and Enterprise, custom priced with SSO, RBAC, and a dedicated VPC.

What FlowRunner adds around it

Look at what CrewAI’s human feedback hooks hand back to you. The async provider has to notify someone and the code that resumes has to be written. On AMP, the webhook arrives at a URL you own. Who gets the request, in which app, what they click, and what happens if they take three days: that’s your code. CrewAI Enterprise adds HITL management for responder assignment and escalation. On the open source framework and AMP Basic, you build it.

FlowRunner is a workflow platform with that layer already in it. A flow is a set of blocks on a canvas, and the parts a crew needs in production are blocks you don’t have to write:

• A review that reaches people where they work. A flow sends the request by email, Slack, WhatsApp, phone, or a Form with approve and reject choices, then pauses until the answer comes back. The reviewer doesn’t need a FlowRunner login or a screen you built.
• Long waits. A run can wait up to 30 days on Free, Starter, and Growth, and up to a year on Professional and above. The whole run, wait included, counts as one execution.
• Triggers from 2,100+ integrations. A new row in a sheet, a deal moving stage, an email arriving, a webhook from any system, or a schedule (once, every few minutes, daily, weekly, monthly, or cron) can start the flow that kicks off your crew.
• Retries and error paths. The HTTP Request block that calls your crew retries on the status codes you list (429, 5XX) with fixed or exponential backoff, and a Handle Error block gives a failed step a recovery path instead of a dead run.
• A record of every run. Each run is on the flow’s Instances tab with the input and output of every step, including the review and who answered. Anyone in the workspace can open it.
• Unlimited users on every plan. The ops lead who approves, the manager who audits, and the developer who maintains the crew all get access, at no extra cost per seat.

None of this replaces anything in the crew. Your agents, tasks, and tools stay in Python. FlowRunner runs the parts around them.

How they work together

These are the patterns that come up most. The code for each is on Use CrewAI with FlowRunner.

Your Flow asks a person, FlowRunner gets the answer. Write a HumanFeedbackProvider whose request_feedback posts the method’s output to a FlowRunner flow through the non-blocking Call Flow API and then raises HumanFeedbackPending. The flow posts the draft to the reviewer in Slack (or sends it by email or a Form), waits, and when the answer comes in, calls your small resume endpoint. That endpoint loads the paused Flow with from_pending and calls resume() with the feedback. If you use emit, CrewAI’s own outcome routing takes it from there.

AMP’s human input webhook lands in a FlowRunner flow. On CrewAI AMP, set humanInputWebhook at kickoff to the Callback URL of a FlowRunner flow that starts with an External Callback block. The flow gets the execution and task ids and the task output, reaches the reviewer, and calls AMP’s /resume endpoint with human_feedback and is_approve. If the reviewer rejects, AMP has the crew retry with their comments, exactly as CrewAI documents.

A business event starts the crew. A flow starts on any of FlowRunner’s integrations or on a schedule and calls your crew: AMP’s /kickoff endpoint, or your own service that calls kickoff(). The call sits in an HTTP Request block with a Retry Policy, and the run record shows what started the crew and what it was sent. AMP’s own triggers still work. This covers the apps and webhooks outside that list.

The crew uses flows as tools. Wrap the blocking Call Flow API in a CrewAI tool. The flow does the work with FlowRunner’s integrations (look up a customer in the CRM, create an invoice, post to a channel) and returns the result in the same request, within up to 300 seconds. The agent gets a typed answer, and each call shows up in FlowRunner’s run history.

FlowRunner Prompt to Flow builds the FlowRunner side. You don’t need to learn the flow editor first. Connect FlowRunner Prompt to Flow to Claude Code, Cursor, Codex, or Claude, describe the review flow in a prompt, and the AI builds it in your workspace and tells you what your code sends and gets back.

When you don’t need FlowRunner

Some crews are fine on their own, and adding a platform to them is overhead:

• You’re the only reviewer and you’re at the terminal. human_input=True or the default console provider does the job.
• The crew runs from your code, on your schedule, and only you look at the results. kickoff() from a cron job on your own server is enough.
• It’s research or a prototype. A notebook that tries out roles and prompts doesn’t need triggers, approvals, or a run record.
• AMP’s triggers and webhooks already cover it. If your crews start from Gmail, Slack, HubSpot, or Salesforce and you’ve built the review receiver you need, there’s nothing missing.
• You’re on CrewAI Enterprise and its HITL management fits your reviewers. Responder assignment, escalation, and SLAs are part of that plan.

Add FlowRunner when the crew becomes part of a business process: someone outside engineering decides, the decision can take days, the work starts in a business app, or other people need to see what happened.

Feature comparison

Dimension	CrewAI on its own	CrewAI with FlowRunner
Agents, tasks, crews, flows	Defined in Python	The same code, unchanged
Human input	human_input, @human_feedback, AMP webhook loop	The same hooks, with a flow on the receiving end
Who reaches the reviewer	Your provider or webhook receiver (Enterprise adds HITL management)	A flow, by email, Slack, WhatsApp, phone, or a Form
How long a review can wait	As long as your persistence keeps the state	The flow waits up to 30 days (Free to Growth) or a year (Professional and up)
What starts a run	kickoff() or crewai run; AMP adds 10 app triggers	Any of 2,100+ integrations, a webhook, or a schedule
Schedules	Not in the open source framework	Once, interval, daily, weekly, monthly, cron
Retries	Task guardrails, guardrail_max_retries defaults to 3	Plus Retry Policy and Handle Error around every call to the crew
Run record	AMP traces	Every run’s steps, inputs, and outputs, open to the whole workspace
Users	Per your AMP plan	Unlimited on every plan
Building the integration side	Your code	A prompt to FlowRunner Prompt to Flow, or the flow editor
Price	Framework free (MIT); AMP Basic free with 50 executions a month; Enterprise custom	FlowRunner Free covers 100 executions a month; one execution per run