Level 4 · Systems & agents

How to Hire Your First AI Agent

Not a prompt. A hire. Job description, probation task, feedback. You already know how to do this.

Here's the reframe that makes agents click: you're not learning a tool. You're hiring. And you've been on one side or the other of hiring your whole career.

Quick definitions so we're together: a chatbot answers when you ask. A helper runs a small task on a schedule. An agent is the next rung: you hand it a whole job ("go through this folder and produce X") and it figures out the steps itself: reads, works, checks, and comes back with something finished. Claude Cowork is the friendliest place to do this, so that's where this walkthrough lives.

Step 1: Pick a job you'd give a capable temp

Not your most important work. Not your most sensitive work. A real job with a clear finish line that's been sitting on your someday list:

The receipts pile. "Here's a folder of receipt photos and PDFs. Make me a spreadsheet: date, vendor, amount, category."
The content archive. "Here are 40 old captions. Find the 10 best performers' patterns and give me a repurposing list."
The messy downloads folder. "Sort this by client and file type, flag duplicates, don't delete anything."

Notice what these share: a pile that exists, an output you can check, and no judgment calls that could hurt you if the agent gets one wrong.

Step 2: Write the job description, not instructions

New hires don't get a script of keystrokes. They get: here's the goal, here's what done looks like, here are the rules. Same shape:

The job description shape
"Your job: [the goal]. Done looks like: [the output, specifically]. Rules: work only inside this folder, never delete anything, and if something is unclear or looks off, stop and ask me instead of guessing. When you finish, tell me what you did and anything you weren't sure about."

Those last two rules are the whole safety story: ask instead of guess, and report what you did. Every good agent setup has them, and honestly, every good manager gives them.

Step 3: Watch the first run like it's a working interview

Stay at your desk the first time. Watch what it does, answer its questions. This isn't wasted time, it's exactly what you'd do with a human's first assignment. You're learning what it's great at (volume, consistency, never getting bored) and where it needs direction (your preferences, your exceptions, the thing only you know).

Step 4: Give feedback like a manager, not a perfectionist

The output will be a B+ the first time. Resist the urge to fix it yourself, that's how you end up with an assistant you don't trust and a job you still do. Instead, say what you'd say to a person: "Category names are too generic, use these instead. And put refunds in their own tab." Then, and this is the move, add that feedback to the job description so the next run starts smarter. Your corrections become the training.

Step 5: Promote slowly

The trust ladder is the same as with people: watch it work → spot-check its work → hand it the routine and check monthly. One rung at a time, one job at a time. When the receipts job runs clean twice, give it a second job. Six months from now you'll have a small staff that costs less than your streaming subscriptions.

And if anyone asks what you do now, you're allowed to say "I manage a small team." You don't have to mention that the team doesn't drink coffee.

Want to see what agents can take off your plate?

The Cowork gallery has real jobs to steal, filterable by the part of your life that's drowning.

Browse the gallery →

When you're ready for more

In the cohort, you hire your first agent with me watching the working interview alongside you. Learn about the cohort →