Agents vs Workflows: What Should You Actually Automate?

Every AI pitch right now ends with the same question. Do you want an agent?

I understand the pull. The word sounds like an employee who never sleeps. But before you buy one, build one or nod along in a demo, it helps to know what the word means, because the people who build these systems for a living use it more carefully than the people selling them.

Anthropic published a guide on December 19, 2024 that draws a clean line. On one side, workflows. On the other, agents. Most of the money a small business will save with AI sits on the workflow side.

What is the difference between a workflow and an agent?

It comes down to who decides the next step.

Anthropic's definitions

In "Building effective agents," Anthropic describes workflows as systems where LLMs and tools are orchestrated through predefined code paths. Agents, by contrast, are systems where LLMs dynamically direct their own processes and tool usage, keeping control over how they accomplish tasks.¹

In plain English

A workflow is a recipe, and recipes start as written procedures (see your first AI hire is a writer). You decide the steps, the order and what happens when something goes wrong, and the model does a piece of the work inside each step. An agent is closer to a contractor you hand a goal. It decides what to do next, which tools to use and when it is done.

The recipe is easier to predict, test and audit. The contractor can handle messier problems, and it can also wander.

Why are most small-business wins still workflows?

Because your business runs on repeatable steps.

Anthropic's own advice

Anthropic recommends finding the simplest solution possible, and only adding complexity when needed. It notes that agentic systems often trade latency and cost for better task performance, and that this tradeoff should be weighed. For many applications, it says, optimizing single LLM calls with retrieval and examples is enough.¹

That is the people who build the tools telling you not to over-build. I would take them at their word.

My read for small operators

You have a handful of repeating jobs: a new lead arrives, a client asks a question, an invoice needs a reminder, a proposal needs a check. Each has a known shape. The value of automation is speed and consistency, and both come from predictability. An open-ended agent adds flexibility you rarely need and unpredictability you cannot afford when the output goes to a customer.

What do the workflow patterns look like in a small business?

Anthropic lists several workflow patterns. Here is how I would translate each.¹

Where this gets interesting

Here is the first pattern as a picture, using a new lead as the example.

Notice that four of the five are about structure you define in advance. The fifth, orchestrator-workers, lets a model decide how to break up the task, which is closer to an agent. That is where I would slow down and add checks.

When does an agent actually make sense?

When the steps cannot be predicted and a mistake is cheap.

Open-ended, low stakes

Anthropic describes agents as suited to open-ended problems where it is hard or impossible to predict the number of steps and where you cannot hardcode a path.¹ Examples it gives include a customer support agent that can pull customer data and order history, and a coding agent whose work can be checked by automated tests, with human review still important.¹ Notice what both have: tools, a way to verify results, and a person watching.

The tools are early

Two pieces of infrastructure appeared recently. In late November Anthropic released the Model Context Protocol, an open standard for connecting AI assistants to the systems where data lives, which is why I wrote about getting your data ready first.² In October it put computer use into public beta, which lets Claude operate a screen, and it describes that capability as at times cumbersome and error-prone (I walked through safe first tests).³ Both are interesting. Neither is a reason to hand an agent the keys to your business this quarter.

On Y Combinator's Lightcone podcast, Jared Friedman made the case for agents in a video posted on November 22, 2024: "the vertical AI equivalent is just going to be the software plus the people in one product."⁴ That is the high-end bet. For a small business, the same logic points to starting with a boring, repetitive admin task and keeping a person in the loop until it is reliable.

Plays here, from YouTube

How do you decide? Four questions.

Start with the dullest job

Pick the task you do most often that follows the same shape each time. Anthropic's advice is to start with simple prompts and add multi-step systems only when simpler solutions fall short.¹ A good first project is the reply to a new lead, with your approval before sending. Run it on your last twenty real inquiries and see what it gets wrong.

What guardrails keep it safe?

Boring ones. They work.

• Human approval on anything that reaches a customer or spends money (the same rule I use for AI voice agents).
• Narrow permissions. Give a system access only to what it needs for one job.
• Logs. Be able to see what the system did and why.
• Real test cases. Use your own messy examples, not just clean demos.
• A kill switch. Know how to turn it off in one step.

I run my business alone, so I want every hour back that I can get. That is exactly why I build the predictable thing first and let the clever thing earn its way in. If you want help deciding what to automate first, that is the work I do.

Keep building.Ekai