I Was an AI Agent in College. Here's the Test for Your Job.

The last job I held that had nothing to do with computers was peer advisor at the University of South Florida. This was the mid-1980s. I did not get it for my technical skills. A friend told me about the opening and encouraged me to apply. The woman who hired me chose me over that friend. She thought I would be more relatable.

I have been thinking about that job a lot lately. Of every job I have held, including all the ones since with computers in them, it was the most robotic.

What the Job Was

I worked in the Division of Undergraduate Studies. That is where new students were placed before a college accepted them: Engineering, Business Administration, Nursing. My job was to advise fellow students on which classes to take to move toward declaring a major.

The guidelines were tightly specified. I had very little leeway, and that was a good thing. They were guardrails.

Nursing and Education had the strictest requirements, stricter even than Engineering, which is where I ended up. The easiest majors lived in the College of Arts and Letters: English, History, Philosophy. Although I can't remember ever advising a philosopher or an economist. Interesting. Plenty of Mass Communications majors, though. Today that would be a major in TikTok with a minor in Instagram Modeling.

Why She Hired Me

It is an odd fit. The one job I ever had that ran on rules was the one where I was picked for people skills.

The rulebook did not need my personality. It needed someone a new student would be willing to sit down across from. That was the hire: a friendly face on a fixed process. The advice itself came out of the guidelines. I did not think of it that way then. I do now. I was the chat window.

Guardrails Are Everywhere

Humans work inside guardrails all the time. The AP Stylebook. The DSM-5. ISO 9001 certification. Each one exists to make the product uniform, so the rules do not have to be worked out again at every session. A copy editor does not rebuild the rules of grammar for every story. A clinician does not rebuild the diagnostic criteria for every patient.

A system prompt does the same job for an AI agent. It is the rulebook written down once, so the model does not have to guess at every session.

The Anatomy of My Job

Look at how the job was built. Each student arrived with a context: a transcript, some vague idea of what they wanted to do with their life, and a set of abilities they had spent eighteen years acquiring. The institution above me supplied what an AI engineer would call the system prompt. I supplied my limited knowledge of how it all fit together, from the half day of training I received. And I had a template to fill out: a little scorecard listing the courses they should take next semester.

Context. System prompt. Light training. A form to fill out. Hand a stranger that description and they would think you were describing an agent.

And I was one, because I did something. Advice on its own is chatbot work. My output was a recommendation form, and filling it out was an action. It was one of those multi-part forms on carbonless paper, where the pressure of the pen makes every copy at once. That form created a record, in copies. One went to the student. One went into the department file. Once it was filled out, it was no longer just something I had said.

An AI agent could have done that job as well as a college sophomore in Engineering. Probably more consistently. I don't say that to insult the job. I say it because it describes many of the jobs we are now deciding whether to automate. It also means the rulebook was doing more than I thought. The rules did not only shape my advice. They shaped what got filed.

The Box That Wasn't There

The scorecard had a line for every course. It had no line for the vague idea. That idea was the most interesting thing in the room, and my job was to fit it to the catalog, not the other way around. Eighteen years of abilities went in one end, and a list of classes came out the other. Uniformity was the product. It was the sausage grinder of conformity the Florida university system was founded to run, and I was a small gear in it.

An agent would run those guardrails faithfully, including the ones that should not exist. A human advisor with a little leeway can notice when a rule does not fit the person across the desk. Mine was small. Take the human out and even that margin goes with them. The rules stay exactly as they were.

What to Do With This

Here is the test. Write down four things about the job: the context it receives, the rules it must follow, the specific action it has to take, and how you will check its work. If you can fill in all four, the job is ready for an agent. If you can't, that is the work to do first.

Read the rules as if you were the person on the other side of the desk. If you would defend every rule, hand the job over and put your effort into the rules and the action. If a rule makes you wince, fix the rule first. An agent will not wince for you.

Then plan for both outcomes, the way my job did. Supervisors reviewed every recommendation before it went anywhere. Once, one of mine was wrong, and I had to call the student back and correct it. That process was as defined as the rest of the job: who catches the error, who gets told, how it gets fixed.

That review step has a limit built into it, and the limit is you. Your eyes and ears take in something on the order of eleven million bits of raw information a second. In 1989, the physiologist Manfred Zimmermann estimated that only about fifty of them reach conscious awareness. An agent has no equivalent ceiling. It will generate work faster than any reviewer can absorb it. I've made the fuller case for this elsewhere: the hardest part of putting an agent to work is rarely the agent. It is budgeting the one resource that cannot be scaled, human attention, the most valuable commodity there is.

I was hired for being relatable, and I spent the job following a script. I never met the person who wrote the scorecard. Whoever it was decided what a finished student looked like. Who wrote the scorecard your agents are filling out?