Flows: chaining steps that don't need a brain

A flow is a sequence of steps that runs on a schedule or a trigger, mostly deterministic, with an AI step or two mixed in. Think of it as plumbing.

Flows vs. agents (again)

A flow is the boring, reliable cousin of an agent. It doesn't decide anything at a high level — you decided everything when you drew the flow. AI shows up as a step: "classify this email," "summarize this doc," "draft this reply." The flow itself is just Python or a visual boxes-and-arrows canvas.

Where you can build flows

Case study: inbound-lead intake flow

Trigger: a new email to hello@your-company.com. Same pattern works for a design agency, a real-estate broker, a plumbing contractor, or a small SaaS.

Build it in n8n (visual)

n8n is a free, open-source visual flow builder. You can self-host it on a Cloudflare tunnel or run it locally. The above flow, in n8n, is roughly:

1. Gmail Trigger node — polls every 5 min for new mail to hello@.
2. OpenAI / Anthropic node — prompt classifies to NEW_LEAD | EXISTING_CLIENT | SPAM, returns JSON.
3. Switch node — routes based on the classification.
4. NEW_LEAD branch: OpenAI node extracts {client, budget, deadline} → Notion node creates a CRM card → OpenAI node drafts a reply → Gmail node saves as draft.
5. EXISTING_CLIENT branch: Gmail node adds a label, sends Slack ping to #projects.
6. SPAM branch: Gmail node moves to archive.

Total time to build once you have n8n running: 30–45 minutes. Total lines of code: zero.

Build it in Cloudflare Workers (code)

Because you already deploy on Cloudflare, here's the same flow as a Worker triggered by Gmail push:

TypeScript · Cloudflare Workerexport default {
  async fetch(request, env, ctx) {
    const email = await request.json();

    // 1. Classify with a cheap model
    const cls = await classify(email, env);
    if (cls === "SPAM") return new Response("ok");

    // 2. Route
    if (cls === "NEW_LEAD") {
      const details = await extractLeadDetails(email, env);
      await createNotionCard(details, env);
      const draft = await draftReply(email, details, env);
      await saveGmailDraft(draft, env);
    } else if (cls === "EXISTING_CLIENT") {
      await labelEmail(email.id, "active-project", env);
      await pingSlack("#projects", email, env);
    }

    return new Response("ok");
  }
};

async function classify(email, env) {
  const res = await fetch("https://api.openai.com/v1/chat/completions", {
    method: "POST",
    headers: {
      "Authorization": `Bearer ${env.OPENAI_API_KEY}`,
      "Content-Type": "application/json"
    },
    body: JSON.stringify({
      model: "gpt-5-mini",
      messages: [
        { role: "system", content: "Classify to NEW_LEAD, EXISTING_CLIENT, or SPAM. Reply one word." },
        { role: "user", content: `Subject: ${email.subject}\n\n${email.body}` }
      ],
      temperature: 0
    })
  });
  const json = await res.json();
  return json.choices[0].message.content.trim();
}

Deploy with wrangler deploy. The whole thing runs at the edge, no server to maintain, roughly $0.30/month at your volume.

The simplest flow: a scheduled task in Computer

You don't need n8n or a Worker for many flows. In Perplexity Computer, you can just say:

That's a real cron job. It writes the schedule, wires the connectors, and runs while you sleep. We'll build one properly in Module 8.