Your AI Agent Is a For-Loop in a Trench Coat
The word "agent" is doing the work the software can't. We renamed automation and raised a round on it.
Sometime in the last year, half the products I use quietly rebranded. The chatbot became an "agent." Nothing under the hood changed. Same model, same API call, same box you type into. But "chatbot" was 2023 money and "agent" is 2026 money, so the word got swapped and the price went up. I want to talk about what that word is hiding.
An agent, in the sense the term got borrowed from, is supposed to do something real. Read a situation, form a plan, take actions in the world, notice when it's wrong, and adjust, all pointed at a goal you handed it and mostly without you standing over its shoulder. That's the pitch on the landing page. An autonomous digital worker. Give it a task, walk away, come back to finished work.
Now open the thing up. What you usually find is a flowchart. Step one calls the model. Step two parses what it said. Step three hits an API. Step four loops back if some condition isn't met yet. There's a language model somewhere in the middle making one or two judgment calls, and wrapped around it is a scaffold of plain, deterministic code that a developer wrote by hand and could draw on a whiteboard in five minutes. We already have a name for that. It's a script. It's a workflow. It's a for-loop with a nice UI.
Remember AutoGPT? Spring of 2023, it trended everywhere. Give it a goal, watch it spawn its own sub-tasks and chase them without you. It was mesmerizing for about a weekend, until people actually ran it and watched it loop forever, burn through real money in API credits, and confidently deliver nothing. That was the honest version of the dream, and it flamed out in public. The current wave learned the lesson, and the lesson was not "make it truly autonomous." The lesson was "hide the unreliable part behind a human and keep the word."
The autonomy they sell you is exactly the autonomy they can't afford to let run.
Here's the tell. Watch where the human is. Every "agent" product that touches anything that matters has an approval step bolted onto it. It drafts the email and waits for you to hit send. It proposes the code change and waits for review. It fills the cart and asks before it buys. Vendors frame this as a thoughtful safety choice, keeping you in control. It isn't. It's there because the fully autonomous version fails often enough that letting it run loose would be a lawsuit. The leash is not a feature. The leash is the confession.
The genuinely autonomous software, the stuff that actually earns the word and runs unattended, is boring. It's narrow. It does one constrained job, retries when it fails, screams into a log when it can't recover, and nobody films a keynote about it. We've had that for decades. We called it a cron job, a daemon, a pipeline. The reason nobody stapled "agent" onto it back then was simple. There was no round to raise by doing so.
I'm not saying the tech is fake. Point one of these at a narrow task where checking the answer is cheap and it's genuinely useful. Sorting tickets. Drafting a first pass. Pulling structured data out of a mess. The trouble starts the second "useful assistant on a short leash" gets sold as "autonomous employee you can cut headcount over." That gap is not a rounding error. That gap is the whole business model, and someone is pricing per "agent" like you're hiring a person, when what you're renting is a subscription to a for-loop.
The word is load-bearing. "Agent" lets a company charge per seat instead of per feature. It lets a deck say "digital workforce" instead of "we automated three steps of your invoice process." It reframes a tool you operate into a colleague you manage, and colleagues justify a much bigger number on the invoice. The linguistic inflation is doing the work the software can't do yet, everyone in the room knows it, and nobody says it out loud because the valuation leans on the word.
So here's a test you can run on any "agent" you're about to pay for. Ask what happens if you pull the human out of the loop entirely and let it run against real stakes for a week. If the honest answer is "we'd never do that," you don't have an agent. You have a chatbot with a project manager stapled to it. That can be worth the money. It is not the thing on the label.
Call it what it is. Most of what's being sold as an autonomous agent today is automation wearing a costume, and the costume costs extra. The software will get there eventually. The word already showed up, alone, way ahead of the thing it was supposed to describe.