What I'm learning putting AI to work inside real organizations.

Automation, workflow, agent: stop calling all of it agentic

Daniel Cornelius · November 10, 2026 · 4 min read

A builder spends an afternoon on an agent. They write careful instructions, attach a few tools, and test it. The job: take a name and an address, and fill a CRM field with a URL.

It works. Then I ask what the URL depends on. Only the name and the address. Same input, same output, every time.

That isn't a job for an agent. It's a formula. The builder paid for reasoning they didn't need and got a little variance in return.

This keeps happening because the vocabulary is blurry. "Agentic workflow" now gets applied to nearly anything with a language model in it, so the term no longer tells you what you're looking at.

Why the term got slippery

In 2024, Andrew Ng popularized "agentic workflows" to describe iterative LLM patterns such as reflection, tool use, planning, and multi-agent collaboration (his Snowflake Summit talk, June 2024). Anthropic's Building Effective Agents (December 2024) drew a sharper line: workflows run through predefined code paths, while agents direct their own process and tool use. (Observation: since then, vendors have applied "agentic workflow" to everything in between, so the label rarely says which one you're getting.)

Three families

Who decides the route Where the LLM appears How the route is enforced What varies run to run
Automation A person, in advance Nowhere By code Nothing
Workflow A person, in advance Inside one or more steps By code The content of a step, not the route
Agent The model, at run time Everywhere, including routing By a prompt, or not at all Route and content

Automation is a formula, a script, a rule. If you can write the logic down, it belongs here, and it's cheaper and more reliable than anything with a model in it.

A workflow keeps the route fixed in code and puts a model inside the step that needs judgment: classify this email, summarize this call, draft this reply. Everything around that step stays deterministic.

An agent gets a goal, tools, and knowledge, and chooses the route itself on each run.

PERSON DRAWSMODEL CHOOSESENFORCED BYAutomationNo modelCodeWorkflowModel inside one stepCodeScripted agentYour steps, model drivesPromptcontrols in codeOpen agentModel picks the stepsPromptor nothingReadRuleFormatWriteTriggerClassifyCheckUpdateLook upResearchDraftCheckFixed in codeModel chooses the next stepModel judgment

The useful middle: scripted agents

Most real business agents sit between the last two rows. A person writes the procedure in a prompt, step by step, with tools at each step, and the model decides how to move through it and which rules apply. The route is designed in advance, but it's executed by the model.

This is a strong fit when the work has many possible routes, or when the inputs are messy enough that hand-coding every branch isn't worth it. In my own work I've built a number of go-to-market processes this way: researching an account and its contacts, then preparing outreach for the rep who owns them, and triaging inbound partner cases with many possible routes. (Observation from my own work, not a measured result.)

But it matters who enforces the route. In a workflow, code does: step three can't run before step two. In a scripted agent, the prompt does, and a prompt is a request, not a guarantee. The model will usually follow it. Nothing makes it.

My rule of thumb is a hard skeleton with a soft middle. Anything with one right answer, like who owns a record, which field to write, or which check is required, should be a lookup or a gated step. Leave the judgment to the model. And the controls (permissions, required checks, approval limits) belong in code and in tool permissions, never only in the prompt.

This isn't a ladder

It's tempting to read the table as a maturity scale, with agents at the top. It isn't. The right design is whichever matches the shape of the work. A lookup is automation. A messy request with one fuzzy decision is a workflow. A case with many routes is a scripted agent. An investigation across inputs you can't predict is an open agent. Most real systems are a mix.

Four questions to ask when someone says "agentic workflow"

  1. Does any step need judgment over messy input? If not, it's automation, and the model is cost without benefit.
  2. Who decides the next step: your code or the model? Code means workflow. The model means agent.
  3. Is the route enforced by code, or just described in a prompt? If it's a prompt, you have a scripted agent, and your controls need to live somewhere harder.
  4. If I run it twice on the same input, does the route change? If only the wording varies, it's a workflow. If the route itself changes, it's an agent.

Then ask the one that matters most: what does a wrong route cost? The higher the cost, the firmer the controls need to be, whichever design you picked. That's the blast radius from last week, applied to the path.

These questions are polite enough to ask in a meeting. If the answer to the first is "no," you've just saved someone an afternoon.

What's next

You now have the vocabulary. Next week: how to tell whether the design you picked is actually working. A scorecard.

Where have you seen an agent doing a job a formula could do, or the reverse? I'd like to hear which one cost more.

Enjoyed this? Subscribe by email on Substack or via RSS. I post one note a week.

Discuss this post on Substack →