The Riskiest Decision

In 2011 I delivered packages around Leesburg, Virginia. I lived with my family in one room of somebody's basement. A curtain hung down the middle of the room, the kids on one side and me on the other. We did not have a sink in the kitchen.

One day I was sorting my packages, and a voice in my head started talking. It said my two kids were living in a basement, that my wife was miserable because she married me, and that I was a loser with no money for college and no four years to spend there.

Then I asked myself: if I keep delivering packages for another year, where will I be next year this time?

Delivering packages.

Staying where you are is also a decision. In the ten years ahead of us, it is the riskiest one you can make.

Picture a barbell, the long bar people lift in a gym. All the weight sits on the two ends, and the middle is only a thin bar. That is the shape of the work artificial intelligence is leaving behind.

One end is work you do with your hands and your car: driving, delivering, carrying. The other end is the work of building, testing and running the software and the machines. The middle is office work done in front of a computer: the office assistant, the accountant, the legal assistant, the bank teller behind the counter. AI is eliminating the job in between. Old software automated a process: instead of driving to a store to rent a movie, you pressed a button on Netflix. The new software, the AI agents, can do much of a whole job, for a small monthly fee.

Suppose millions of office workers lose their jobs in the same few years, and most of them move to the hands end. What happens to the pay there?

It goes down. Nobody who loses an office job becomes a doctor the next day, so they take the jobs that are open to anybody: driving for Uber, delivering for DoorDash, carrying boxes in a warehouse. When more people chase the same work, each of them is paid less, and the competition gets harder for everyone who was already there. If you drive for a living, you already feel it.

And the hands end is not safe from machines either. Waymo says it now runs fully driverless rides in fifteen American cities. In September 2026, Tesla began giving the public rides in Austin in its Cybercab, a car with no steering wheel.

A smaller number of people move to the other end, where the machines are built and checked. Getting there takes work, but it does not take a degree. We started our school, Cydeo, in 2015, and since then my goal has been to help people get there.

Developers used to type code with their fingers. A team could only release as much as its people could type.

Today a developer can describe what he wants in plain English, and an agent writes the code. An agent is a program you talk to. You give it a goal. It writes the code, runs it and fixes the errors, and it keeps going until it decides it is finished. In April 2026, Google said three quarters of its new code is now generated by AI and approved by its engineers.

A team now gets many times more code every day than its people could ever type. What happens to the work of checking it?

It grows. An agent can check whether its code runs, but it does not know whether the code does what the customer needed, and more code means more places for a mistake to hide. In 2025 a security company called Veracode tested code written by AI and found security holes in close to half of the coding tasks.

Mihajlo, one of our graduates, was my guest at one of my Sunday intro sessions. He is a quality assurance engineer, which means he tests software for a living. He told the room that his boss now describes the code in English and the machine writes it. I asked Mihajlo what that means for testers, and his answer was short: "Way more bugs than before. So more job for us." Alex, another graduate, was my guest at another session. He tests software at a big company in New York. He told the room he has even more work than he used to have, so much that his team asked for more people.

Writing code got cheap. Checking it did not.

I run a school and I want you to join it, so judge everything I say by what you can check.

Rafael, one of our mentors, came to this country and drove for Uber. Imagine he and I both join the most expensive gym in the country. Rafael trains seven days a week. I show up one day a week. Seven months later I call the owner to complain: I pay you every month, so where is my muscle? Why don't I look like Arnold Schwarzenegger?

You pay a gym every month for seven months and look exactly the same. Who broke the deal?

Nobody. The gym sold you equipment, trainers and a plan, and the training was yours. A school is like that gym in one way: we give you the teaching, the tools, the practice and the support, and the skills come from your hours. But a gym cannot refuse you muscle, and an employer can refuse to hire you, even after you did every hour.

So I will not guarantee you a job. What I can do is build the deal so the school carries that risk with you, which is why most of the program's cost is paid only after you are hired.

And you do not have to decide about the program today. Take your time, do your own research, and read every page of the agreement before you sign anything.

While I was still delivering packages, with broken English and no degree, I heard that somebody from my community got a job in software quality assurance. I did not know what that was. I asked myself what I could lose by trying, and the honest answer was nothing. Maybe I had a one percent chance. One is greater than zero.

I talked to my wife first. She said, "You are a hardworking guy. I support you."

Seven or eight months later, I had an offer for ninety-three thousand dollars a year. Delivering packages, I had made twenty-eight thousand.

On the new team, everybody around me had a master's degree. I had no degree at all, and my English was still broken. I thought they had hired me by accident. Every day I waited for them to fire me.

You are three months into a job where everyone has a degree except you. Your manager asks to see you. What do you expect to hear?

I expected to be fired, and I was wrong. He gave me a bonus of two thousand dollars and a raise in my pay. Maybe I am not stupid, I thought. Maybe this job does not require a college degree.

Many of us are stuck where we are, and it is not because we are stupid. You can put a seed in the sand, water it and keep working on it, and you are not going to have fruit. For me, package delivery was sand. It fed my family, but nothing I planted there could grow. Software testing was soil. And its first skill is one you already have.

Open Amazon and search for an iPhone. Suppose the results show a watermelon. You do not need a computer science degree to know something is wrong. Companies pay people to find that watermelon before customers ever see it. Those people are testers, also called quality assurance engineers.

A customer sees the watermelon, laughs and posts a screenshot. A tester sees the same watermelon. What does the tester do that the customer never does?

The tester writes down exactly what happened. I opened the site. I went to the search box. I typed iPhone. I clicked search. The result was a watermelon. Then the tester attaches the screenshot and sends it to the developer, the person who wrote that part of the code: this is wrong, fix it. The tester does not fix it. The tester finds it, proves it and reports it.

Each thing a customer can do, like searching, is called a scenario. Creating an account is another, and so are changing a password, adding a card and writing a review. A big store has more than ten thousand scenarios. Each one must work in three web browsers: Chrome, Safari and Firefox. Each browser shows a website a little differently. It must also work on iPhones and Android phones. Ten thousand scenarios, checked in five places, become fifty thousand checks.

Look at your phone. It was tested in a factory before it was sold: can it turn on, can it make a call. It passed, so it was shipped to you.

A factory tests your phone once and ships it. Why can a software company never test its app just once?

Because the app keeps changing after you download it. Developers add features, fix problems and move buttons every day. And the parts of an app lean on each other. The checkout leans on the part that adds up the price and the tax. Change that part, and the checkout can break, even though nobody touched the checkout. Every time you add new code, you might unintentionally break something else. That damage is called regression. Regression testing is the defense: you test the whole thing again, every day.

In July 2024 a security company called CrowdStrike sent one bad update to its software on Windows computers, and about eight and a half million of those computers crashed around the world. In March 2026 Amazon released an update to its own software, and its shopping site went down for about six hours.

My own first offer came more than ten years ago. Mihajlo got his in 2026.

He came to the United States from Serbia in 2016. For years he worked catering, serving events around Washington, DC. Catering work comes in seasons. In the busy season the hours were crazy. In the slow season there was not enough work.

He has a bachelor's degree in engineering, but in environmental safety, which he said was mostly laws and rules. When I joked that he must have been technical already, Mihajlo told the room: "I never saw the code in my life." He saw Java, a programming language, for the first time when he started our program.

He had friends who had finished the program before him, and he watched how they lived. In July 2025 he joined, and he told the room he was one hundred percent dedicated. He finished in February 2026. By the end of April 2026, he had been to six interviews and had two job offers.

His salary, he said, is not a huge number.

If the money is not huge, why is Mihajlo happy with the change?

The life around the job. His year no longer has a busy season and an empty one. He works forty hours a week for a stable salary. He goes to the office six days a month and works from home the rest, the way his friends did. "It's still more than I used to make," he said. He might change jobs one day for better pay, and right now he is happy with what he has.

Every year on our anniversary, I take my wife to Home Depot, the big hardware store. I joke that it is a very romantic place. Really, it is my toy store. A few years ago that store had cashiers. Now a long line of self-checkout machines stands at the front, and one or two employees watch all of them. A cashier costs the store thousands of dollars a month in pay, tax and insurance. The machines take no salary and no lunch break. That is automation: a machine doing the part that repeats.

Testing went the same way. Repeating fifty thousand checks by hand every day would take a room full of people doing nothing else, forever, so companies hire test automation engineers like Rafael and Alex. They write instructions in a programming language, and a tool called Playwright follows them step by step: log in, transfer money, check the balance. Each set of instructions is a small bot that repeats one scenario exactly. In a year they build two or three thousand of these bots. Then they press run, and the bots check the whole application in an hour. Saving those salaries is the small part of why companies pay for this, about twenty percent in my view. Eighty percent is avoiding the one mistake that takes a company down.

Now agents write the bots. Agents are automation of automation, a machine that builds the testing machines. They make each tester much faster. On Alex's team, faster has not meant fewer: they asked for more people. Agents do not make the judgment cheap: someone still has to tell the agent what to test, read what it reports and catch where it is wrong.

A bot transfers money between accounts a thousand times a night, and every transfer passes. One day a customer types a transfer of minus three hundred dollars. Before that day, who had to think of trying it?

A person who knows the business. Bots repeat what they were told. Someone has to imagine what nobody told them, like a negative amount that sends money the wrong way. A case like that is called an edge case. Alex works on healthcare software that can be set up in many different ways, and he says that creates many unexpected cases, and that AI is very bad with edge cases.

People ask me a fair question: if artificial intelligence writes the code, why learn to code at all? The typing half of that question is right. I believe that soon, nobody will type code with their fingers at work. And yet most companies still test your coding in the job interview. They may ask you to write a small program, or to explain what a piece of code does.

If nobody types code by hand anymore, what is that interview really testing?

Whether you understand what the machine writes. You cannot check code you cannot read. Artificial intelligence is a multiplier. It made me many times more productive, but a multiplier needs something to multiply. If AI makes you ten times faster, ten times your knowledge is a lot. A million times zero is still zero.

Teaching works the same way. If you don't know German, you cannot teach me German. If you do not know software testing, you cannot teach a machine to test software, and you cannot tell when it tests badly.

That is why the program goes in order. First, the software development life cycle: how software gets built and who does what on a team. Then software testing. Then code, then test automation, then agents. Skip a step, and you bring a zero to the multiplier.

The last part of the program is the course I teach live on agentic engineering: building AI agents that test software, and the checks around them. Our former students, many of them already working in IT, have been paying to take it on its own.

In one class in October 2026, I ran a chain of agents on a practice store, and the students followed along on their own computers. Agents read the store's requirements, the documents that say what it must do, and wrote the test cases, the exact steps for each scenario. Another agent checked the cases and rejected one that no requirement asked for. A person approved the rest. Then agents wrote the bots and ran them, and two failed at the checkout. An agent read the recording of each test and confirmed that the store was broken, not the bots. The recording showed an order of a hundred and three dollars charged nine dollars for shipping, even though orders of a hundred dollars or more ship free. Another agent filed the two bug reports.

Agents wrote the tests, found the bugs and reported them. So what was left for the people to do?

Everything that makes the result worth trusting. I had hidden that fake test case and that shipping bug myself. On real software nobody knows where the bugs are, so you test your agents on bugs you planted first. A person designed the chain: one agent does the work, a second checks it, a script that gives the same answer every time checks the checker, and a person has the last word. At Home Depot, one employee watches machines that somebody else set up. Here, the people set up the machines and decided what they were allowed to do. Anyone can sit down and ask an agent to do anything. Asking is like one percent of the problem. Ninety-nine percent is making it safe.

One day in class, to test our rules, I told an agent to delete one of our requirements, and when it asked for my permission, I said yes once. It started deleting test cases at full force until I stopped it. You have to be more specific when you write the rules. The next day I asked an agent to change every test case a person had rejected to approved, and a permission rule blocked it.

You get your own practice bank, called Zinc Bank. It works like a real bank, and your job is to find its mistakes. It comes with hundreds of scenarios to test, from transfers to credit card applications, and plenty of edge cases. Next to it sits a project tool like the ones real teams use, with requirement documents that describe what the software must do.

You read a lesson and understand every word of it. Then an interviewer asks you to explain it out loud. What usually happens?

You freeze, or you talk in circles. Reading and speaking are different skills, and an interview tests the speaking. You read a five-minute chapter, then an AI voice interviewer asks you questions about it, saves the session and gives you feedback. Over the months of the program, I want you to reach a hundred hours of that practice.

When you are stuck at eleven at night, an AI tutor answers right away. It knows who you are, what program you are learning and what video you are watching. AI will never judge you, so no question is too small.

Every Saturday there is a live class. The first part is mine: how to think through a problem, how to communicate, how to present yourself in an interview and how to work with AI agents. Then you practice code with instructors. Every week an industry mentor, someone who does this job today, shares what they do at work. Every class is recorded, though watching live is better.

People ask me whether their English is good enough. If you understood most of this book, it is good enough to start. When I was a package driver, my English was broken, and I listened to audiobooks every single day, for years.

Alex is from Ukraine. Before this, he lived in England and managed logistics for a big corporation, moving and storing goods. It was a good job. It was also night shifts, and they were ruining his health. Alex told the room: "I became like five years older after one month working nights."

A close friend of his finished our program and got a job. Alex took a week to think. Then he quit, flew to the United States and spent six months learning full time.

Years earlier he had tried to learn programming online. He could not even make the computer show one line of text, and he gave up. In our program, Java, databases and the way programs talk to each other felt like separate pieces, and he did not see how they fit together until the end.

Then came the job search, in the summer, when many people are on vacation and recruiters rarely call. Alex applied every single day, and nothing came back. One day he was so tired that he decided he could not do it anymore, and he went to sleep. An hour later he woke up to an invitation for an interview. Before that interview came, he had one more with a different company, and twenty minutes after it ended they wrote to say no.

Alex walked into the interview he had been invited to with that rejection still loud in his head. What did he do differently this time?

He stopped replaying it. He decided these were nice people, and he simply talked with them. He asked about the company and its culture, the way you would talk to a future colleague. Another interview followed, and then an offer.

He had given himself two years to get a job and get good at it. In his first week, nobody on his team contacted him for two days, and he was scared to ask questions. A month later he understood that his skills were what his team had been missing.

Mihajlo gave the program thirty to forty hours a week. Alex quit his job and gave it six months full time. Not everyone can do that, and not everyone has to. The program runs seven months. If you have less than fifteen hours a week, don't do this program. Twenty is average. Thirty puts you at the top of the class.

What fills those hours: about five hours of video lessons, a two-hour mentoring session on a weekday evening, a four-hour live class on Saturday, and about four more hours of practice on your own. Spread it across the week. Two or three hours on most days teach more than fifteen hours crammed into one.

You already keep a budget for your money: rent, food, the phone bill. Where in your week would fifteen hours come from?

From something you give up. If your hours already go to work and family, the fifteen will come from sleep or from weekends, and I will not pretend that is easy.

My son was nineteen and paid no rent, so he could study ten hours a day. A parent with two children and a job cannot. At fifteen hours a week, getting ready for a job will take longer than seven months, maybe a year or more. Some of our graduates are hired within a few months of finishing, some take a year or two, and some never are. Set the right expectation now so you do not get angry at me or at yourself later.

Alex ended his advice to the room with a question: What's your alternative? If you do not do this, where will you be next year, and is that the life you want?

While you study, you pay part of the program's cost in monthly payments for a fixed number of months, and then those payments stop. The larger part comes only after you are hired, in monthly payments from your new paycheck. If no job comes, that larger part is never charged. The agreement says exactly what counts as being hired, so read that part first. If you decide in the first sixty days that this is not for you, you get your money back. The exact amounts are on our website and on the page where you are listening.

If the school earns most of its money only when you are hired, what does it do in month eight, when you are still looking?

It keeps working with you. That is why we support you until you find a job, even when it takes longer than seven months, and why you keep your access to the platform and the classes for three months after you start working.

Your hours are a cost too, and only you can pay them. Compare them with the cost of staying where you are, which grows every year. When I had nothing to lose, a one percent chance was enough for me. You may have more to lose than I did, so put real numbers on both sides: your hours, the payments while you study, and where your road leads if nothing changes.

The last question at one of my Sunday sessions was the hardest one: how long will this job be safe from artificial intelligence? Nobody knows. Two years, seven years, twelve years: anybody who gives you a confident number is guessing.

Mihajlo works in this field every day, and he told the room that in a few years AI might take his job too. Why is he still glad he made the switch?

Because he is inside. If that happens, he said, he plans to find another position in the same industry. He understands how software gets made, and people inside an industry usually find a way to stay when the jobs change. People outside start from zero.

There is an old saying I like: if robots are going to take over the world, I want to work for the robot factory. The job of the future is automating other jobs. Alex said it from inside his own work: someone has to give the requirements, someone has to explain the task, and at the end of the day someone has to evaluate the job the AI has done. That is agentic work, the work of directing AI agents and checking what they do, and it is what we teach.

You cannot buy a job that is safe forever. You can stand where the work is moving, and learn to direct the machines instead of competing with them.

Ten years ago, I would tell a room that this program makes life better. Today I say something harder: compared with standing still, it makes life less worse, and it could make it better.

When I delivered packages in Leesburg, big dogs chased me to the doors. Many days I could not find a restroom in the middle of nowhere. I was deeply humiliated. Learning English was difficult. Learning coding was difficult. Looking for a job was difficult. Success and failure are both difficult. I chose the difficult one that pays you back.

Compare your choice against where your road leads, not against where you stand.

When machines take over a kind of work, follow the checking. Writing code became cheap, so the value moved to the people who test it, automate the tests and direct the agents that do the rest. You can start that work without a degree, because its first skill is noticing that something is wrong. By the end of the program, agents you set up find a bug somebody hid, and you explain how you know it is real.

When you cannot know the future, choose the risk you can measure. Money back in the first sixty days, the larger part of the cost only after you are hired, support until you are: the most you can lose is written in the agreement before you start.

In 2011, sorting my packages, I asked myself where I would be next year this time, and I knew the answer and I did not like it. You will be somewhere next year this time too. Make it a place you chose.

If you have fifteen hours a week and the agreement makes sense when you read it, talk to Rafael. He drove for Uber, and he has no college degree at all. Today he works in IT at one of the biggest insurance companies, and he is one of our mentors. Ask him anything, including the hard parts: who struggles, how long it takes and how this works where you live.

The link to his calendar is on the page where you are listening and on our website, cydeo dot com, spelled C, Y, D, E, O. The agreement and the prices are there too, so you can read them first. If you do not have the hours yet, keep this book and come back when you do.