Building marketing agents with Make.com: what we learned

Every article about Make.com is either a tutorial by someone selling a course or a competitor’s comparison page. The boring version is the useful one. I spent my evenings this summer building marketing agents in Make.com. This post is what held up, what fell over, and what the credit meter actually does to your bill. Everything time-sensitive is dated to August 2026.

What Make.com actually gives you for agent work

Make is a visual scenario builder, and its agent story is narrower than the marketing says. You draw the flow on a canvas, and the AI agent is a module there, not a separate product. It gets a prompt, a model, and tools it can call. The whole thing runs on the same canvas as your deterministic steps.

The documented guidance matters more than the feature list. Make’s own material repeats the rule (Make blog): scenarios for the deterministic steps, the agent only for judgment calls. One community recap of a first agent build lands on the same point (Make Community). That pattern is the whole post. The canvas is why you pick Make, and it is also where the costs hide.

Two things make Make different from the n8n run and the MCP servers run on this blog. First, webhooks fire instantly, where n8n polled. Second, everything you run costs a credit, and AI modules cost more than one. That second point shaped every decision below.

The build that worked (and the one that didn’t)

The pattern that held up was three small agents, each with one job, on the same canvas. Same lesson as the rest of this series: small specialized agents beat one super-agent.

The research agent takes a topic, searches, and pulls the top sources. It writes a one-page brief with the claims and the gaps, then stops there. The drafting agent takes that brief, writes the draft, and stops; it never publishes, a human reviews every generation. The alert agent watches a campaign metric and messages me when something moves. That replaced a morning ritual of opening three tabs.

The failed version was my first attempt: one agent with every tool attached. It was faster to build and worse at everything, and it burned credits deciding which tool to call. The output needed as much editing as a draft written from scratch. Splitting it into three small agents fixed both. Quality jumped, and credit burn dropped, because each agent only touches the tools its job needs.

The reuse pattern is real too. Make’s sub-agents let you build a specialist once and call it from multiple orchestrators (Make Community). The research agent became a sub-agent that the draft orchestrator and the weekly report both call.

Where Make.com fell over for us

The boring parts broke first, and Make has its own flavor of each.

The credit meter is the hidden tax. Make switched from operations to credits in August 2025, and the difference matters for AI work (Flowbuilder). Standard module actions cost one credit each. AI modules consume credits based on model and tokens, which you cannot predict before you run. My first drafting agent burned through the free tier in an afternoon. The fix was boring: shorter prompts, fewer tools, and a webhook trigger where the source allowed one.

Memory is not built in. An agent in Make is stateless unless you give it state. The community answer, and the one I landed on, is the Data Store (Make Community). You inject past conversation history so the agent can actually use it. It works, and it is another thing to design before you build, not after.

Lower tiers queue. On the Core plan, scenarios can wait in line during peak hours. That does not matter for a nightly report; a lead alert that should fire in seconds feels every delay. The Pro tier adds priority execution, and that is the upgrade reason for anything customer-facing (Make pricing).

Workflows do not port. There is no export from Make that n8n or Zapier will read. Outgrowing the platform means rebuilding on the new one, so budget that into the decision. Or pick the platform you will still be on in two years, because the rebuild is the real cost.

What a Make agent stack actually costs

Entry is cheap, and the bill grows with executions and AI tokens. One “Run an agent” operation costs a single credit (Make Community). The model and token cost of the calls inside it lands on top. That is the line most pricing guides bury.

The thing to watch is not the platform price. It is the credits. A scenario that polls every five minutes runs about 8,640 times a month on the trigger alone. Switch to a webhook and that number drops to however many times the source actually changes (Alltomate). For a solo stack, Core at $12 covers a lot. The moment you add token-heavy AI agents on hourly schedules, model costs become the real bill.

Build or don’t build: the Make readiness check

Build an agent on Make when the task is boring, repeatable, and low-risk to get wrong. Skip it when the work needs judgment you cannot review, touches customer money, or runs once.

• The task happens at least weekly, and the input is predictable.
• A wrong output is caught by review, not shipped.
• You can name the trigger, the tools, and the human gate.
• The credit and token cost per run is a rounding error in the budget.
• You have an error path, so failure is loud, not silent.

That last one matters most in Make: a scenario that fails silently is worse than none, because you trust it. Wire the error handler before you wire the happy path.

And remember the honest test I ran on ChatGPT for marketing: the tool is never the whole story. Make’s canvas is genuinely good at holding multi-step logic that a non-engineer can read. That is its edge over n8n, and its cost edge over Zapier at any real complexity. The agent module is a way to put judgment inside that canvas, not a magic product. If the task is deterministic, a scenario beats an agent every time. Marketing automation agents earn their credits on the steps that need a decision.

Filed under agents; more in the agents category and about this blog.