An AI agent workflow builder built from the routines you already have.

An AI agent workflow is a chain of AI steps where one step’s output becomes the next step’s input, and the whole chain runs on a schedule instead of waiting for you to press a button. In Routines each step is one of your existing routines, so the chain inherits the prompts, connectors, and memory those routines already use.

A multi-agent workflow splits a job across several agents, each with its own instructions, and passes the result along the chain. A flow is exactly that: one routine gathers the data, another interprets it, another decides, another writes. Each one is a separate routine with its own prompt, so you can fix one without touching the rest.

Two ways. They do the work inside a step, like reading your analytics or drafting the article. And they decide whether the next step should run at all: a condition on a connection is a small model call that reads the previous step’s result and answers the plain-English question you typed.

In n8n, Zapier, or Make you wire generic API blocks together, so you own the endpoints, the auth, and the field mapping. In Flows you are chaining your own AI routines, which already have your context, your connectors, and your memory. There is nothing to map, and the branch logic is a sentence rather than an expression.

This one does not. The graph, the schedule, and the run history live on your Mac, and the schedule fires locally. When a step calls a model, that request goes from your Mac straight to your provider using your own key.

No. Flows run on this Mac only. They are not synced to your other devices. If you want the same chain on a second machine, you build it there.

A condition is a model call, not a rule engine, so it has three possible answers: met, not met, and could not tell. You decide what undecided means for each connection with the “If the check can’t decide, stop here” setting. Leave it off and the flow continues. Either way the verdict and the model’s one-line reason are recorded on the run.

The check reads the first ~12,000 characters of the previous step’s result and nothing else. Keep the question about something that appears near the top of that result, and write it as one clear sentence, like “Only if at least one page is losing organic traffic.”

Yes. A step is a reference to the routine, not a copy of it, so one routine can appear in several flows and still keep its own independent trigger for standalone runs.

They are small calls, and the run panel shows the total for you instead of hiding it. In the run pictured on this page five steps and one check came to $0.0021 in checks. Every replay reports per-step duration and the cost of the checks for that run.

An agentic workflow is any process where an AI agent takes more than one step on its own: reading a result and deciding what happens next instead of stopping after a single reply. A flow in Routines is a concrete version of that. Each step is one of your routines, the output of one becomes the input to the next, and a plain-English condition decides whether the chain continues.

Regular automation runs the same fixed steps every time. An agentic AI workflow lets a model read what happened in the step before it and decide what to do next, which is what the condition on a Flows connection does before letting the next step run. That is also why a check can come back could not tell instead of being forced into a yes or a no.