Agentic Workflows: 5 Patterns and How to Pick the Right One

"Agentic" has become a label that gets stuck on everything. A chatbot is agentic. A Zap with one AI step is agentic. A system that runs a whole department is agentic. That makes the word useless when you have to decide what to actually build.

So here is how I use it. An agentic workflow is a process you designed, with an AI model doing the parts that need judgment. The path is yours. The decisions inside it are the model's. That one distinction decides how predictable, how expensive and how fixable the result will be.

What is an agentic workflow?

An agentic workflow is a process where an AI model makes decisions or takes actions at specific steps, while a person has defined the overall sequence of steps.

Compare it with the two things on either side:

Anthropic made this split popular in its engineering guide "Building effective agents", which separates workflows (models and tools on paths defined in code) from agents (models that direct their own process). Their advice is to look for the simplest solution first, and only add complexity when it clearly pays off. I agree with that completely.

If you are still sorting out where plain automation ends and AI begins, my guide to AI automation covers the left side of that table.

Why do agentic workflows matter for a business?

Agentic workflows matter because they let AI handle the messy parts of real processes without giving up the control you need to trust the result.

Most business processes are not open-ended. Invoices, inbound leads, support tickets and weekly reports follow a known shape. What makes them hard to automate is the input: free text, PDFs, voice notes, half-filled forms. A fixed workflow with a model at the right steps handles exactly that.

The market is moving fast in this direction. Gartner expects that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. Yet in a June 2025 forecast, the same firm warned that over 40% of agentic AI projects are likely to get cut by the end of 2027, citing cost, unclear value and weak risk control. The gap between those two numbers is mostly design. Projects that start with a process and a clear structure survive. Projects that start with "let the agent figure it out" tend not to.

McKinsey's 2025 State of AI survey points the same way: only 21% of organizations using generative AI said they had fundamentally redesigned at least some workflows. Designing the workflow is the work. The model is a component.

What are the 5 agentic workflow patterns?

The five agentic workflow patterns are chaining, routing, parallel steps, orchestrator-workers and generate-and-review, and each one fits a different kind of process.

Most of what I see working in real businesses is pattern 1 or 2, sometimes with a review step from pattern 5 on top. That is not a lack of ambition. It is what holds up.

How do you choose the right pattern?

You choose the right pattern by starting from the shape of the process, not from the most advanced option, and by moving up only when the simpler pattern clearly fails.

The questions I ask, in order:

1. Are the steps always the same? Use chaining.
2. Does the input come in a few distinct types that need different handling? Add routing in front.
3. Do you need several independent checks on the same input? Run them as parallel steps.
4. Does the quality bar matter more than speed? Add a generate-and-review loop with a fixed maximum of rounds.
5. Do the subtasks really change from run to run? Only then consider an orchestrator with workers.

Say a 10-person agency gets 40 inbound emails a day. As an example, a routing step sorts them into new leads, client requests, invoices and noise. Leads go to a chain that looks up the company and drafts a reply. Invoices go to the bookkeeping inbox. Client requests get a drafted answer that waits for approval. Noise gets archived. No orchestrator needed. Four routes and one human checkpoint cover it.

If your process needs pattern 4 everywhere, it is often a sign the process itself is not clear yet. Writing it down first usually reveals a fixed structure hiding underneath. My one-page spec for any agent build is built for exactly that.

How do you design an agentic workflow step by step?

You design an agentic workflow by writing the process down, marking which steps need judgment, and adding guardrails and logging before you build anything.

My order:

1. Write the process as it runs today. Every step, every decision, every exception. If a person does it, write down what they look at.
2. Mark the judgment steps. Only those get a model. Everything else stays a plain rule or API call.
3. Pick the pattern. Use the questions above. Start simpler than you think you need.
4. Define inputs and outputs per AI step. A fixed format, ideally structured output, so the next step can check it.
5. Add a human checkpoint where it counts. Anything that goes to a customer, moves money or changes a record people rely on.
6. Give every AI step a fallback. When the model is unsure, it sends the item to a person instead of guessing.
7. Log every run. Input, model output, the path taken, and what the human approved or changed.
8. Run it in shadow mode first. The workflow drafts, a person still does the real work, and you compare.

Steps 5 to 7 are where most builds cut corners, and they are the reason builds die in production. I wrote down the five patterns behind agents that break; every single one is a missing guardrail, not a weak model.

Which tools should you use to build one?

The right tool for an agentic workflow is the one your team can maintain, because the pattern carries the value and the tool only runs it.

For most small companies I'd start visual. You can see every step, and someone other than the builder can follow what happened on a given run. If you go that route, this is how I make an n8n template production-ready.

When should you use an AI agent instead?

Use an AI agent instead of an agentic workflow only when the task has no predictable path, the steps really differ per run, and you can afford stronger monitoring.

Good examples are open research tasks, debugging, or assistants that work across many tools on request. For recurring business processes, a workflow almost always wins. It is cheaper per run, easier to test, and when something goes wrong you can point to the exact step.

In my four-layer view of AI adoption, agentic workflows are the bridge. They take a company from people prompting by hand to processes that run on triggers with a human approving what matters. That is a big step, and it does not need a fully autonomous system to get there.

Pick one process. Write it down. Mark the steps that need judgment. Then build the simplest pattern that covers them.