You probably don't need an AI agent

Someone asks for help with a task. Within a minute, the conversation is about which agent framework to use.

It's the wrong first question almost every time. There are four rungs on this ladder, they get more expensive and less predictable as you climb, and most problems are solved on the first two.

#The four rungs

A prompt is you, a model, and a question. No code. You're in the loop, you see the answer, you decide what to do with it.

A workflow is a fixed sequence of steps that you run. Some steps might call a model; most probably don't. The order never changes, because you wrote the order.

An automation is a workflow with a trigger. Nobody presses the button - a schedule, a webhook or a new file does. It's the same fixed sequence; the only difference is who starts it.

An agent is a loop that decides its own next step. You give it a goal and some tools, and it chooses what to do, looks at the result, and chooses again, until it decides it's done.

That last sentence contains the whole cost. Everything above it takes a path you chose. An agent takes a path it chose, and takes a different one tomorrow.

#What that difference looks like

Both lanes below do the same job. Run them a few times and watch the step counts.

The workflow's honesty is the point: three steps, always three steps, and if it breaks you know exactly which of the three. The agent might finish in seven steps or in thirteen. Sometimes that variability is precisely what you're paying for. Usually you're paying for it by accident.

#Picking a rung

The question isn't "how hard is this task?" - it's "do I know the steps?"

Prompt

You know the steps, and you'll only do it a few times.

Workflow

You know the steps, and you'll do it often enough to be annoyed by doing it by hand.

Automation

You know the steps, and something other than you should start them.

Agent

You genuinely cannot write the steps down, because they depend on what earlier steps find.

That last row is a much higher bar than it sounds. "I don't feel like writing the steps down" is not the same as "the steps cannot be known in advance".

#Four real examples

Summarising a contract. One prompt. You paste the contract, you read the summary, you're done. Building a "contract analysis agent" here buys you nothing except a way to be wrong without noticing.

Turning receipts into a spreadsheet row. A workflow - and note where the AI goes. Read the image, extract the fields, validate them, append the row. Only one of those four steps needs a model; the other three are ordinary code, and they're the three you want to be boringly reliable. The most common mistake I see is handing all four to a model because one of them needed it.

The same thing, when the receipt arrives by email. An automation. Identical steps, different trigger. No new intelligence - just nobody pressing go.

Investigating why a deploy broke. This is agent-shaped. You can't write the steps down, because step two depends on what step one found: the logs might send you to a config change, or a dependency bump, or a bad migration. The branching is the job. This is the rung where the unpredictability is the feature.

Three of those four are solved below the top rung. That ratio matches my experience closely.

#Why the top rung is expensive

Climbing costs you three things people tend to discover late:

1. Debuggability. A failed workflow tells you which step failed. A failed agent gives you a transcript to read.
2. Cost. A fixed sequence has a fixed price. A loop's price depends on how long it decides to think.
3. Trust. Non-determinism is very hard to explain to someone whose work depends on the output. "It usually does the right thing" isn't reassuring, and it shouldn't be.

#Start one rung lower

Try the task as a single prompt, and only build something when that visibly fails. Then write the steps down; if you can, you have a workflow, not an agent. Notice which steps actually need a model - usually one or two - and let ordinary code do the rest.

The best systems I've built have a lot less intelligence in them than people expect. They have it in exactly one place, surrounded by code that does the same thing every time.