AI Agents Won’t Replace You. But They Will Expose You.

Something has shifted in how I work. I spend less time writing code and more time reading it, questioning it, and deciding whether what was just generated actually makes sense. The execution has moved to the agent. The judgment stayed with me.

At first that felt like a productivity gain, and it is. But the more I sit with it, the more I think it's also something bigger. The role of a software engineer is quietly changing, and I don't think everyone has noticed yet.

The conductor and the musicians

The analogy I keep coming back to is an orchestra. A conductor doesn't play every instrument. That's not the job. The job is to hold the full picture, to understand how each part contributes to the whole, and to hear when something is off — even when it's subtle, even when everything technically sounds like music.

That's what building with AI agents feels like. You write less, you direct more. You're not the one typing out every function, but you're entirely responsible for whether the result is actually good. And to do that well, you need to understand the fundamentals more than ever, not less. A conductor who doesn't understand rhythm can't hear when a musician is playing out of time. The abstraction goes up, but the underlying knowledge has to stay solid.

AI won't save you from being a bad developer

This is the part I don't think gets said enough. There's a version of the AI hype narrative that implies the tools are so powerful that the skill floor has dropped — that someone without deep experience can now build things that would have previously required years of expertise. And there's a small truth in that. The tools are genuinely remarkable.

But the floor hasn't dropped as much as people think, it's just moved. The term "vibe coder" had its moment, but the market has been quietly humbling about it. Shipping something that runs is not the same as shipping something that's correct, secure, and maintainable. That gap is becoming harder to ignore.

The gap between a good developer using AI agents and a weak one is arguably wider now than it was before. A weak developer using agents will ship faster, with more confidence, and with less awareness of what's actually wrong. The agent produces plausible-looking output, they trust it, because they don't have the foundation to question it.

A good developer uses the same tools, moves just as fast, and knows exactly where to look.

What a real mistake looks like

While building the authentication flow for Purl, my AI agent wrote the whole thing. It looked clean, it worked, and it passed my initial tests. I was moving fast and feeling good about it. Then I actually read the code.

User passwords were being saved to the database in plain text. No hashing, no encryption, nothing. The row-level security policies weren't properly configured. And subscription data, usage limits, things that have no business living on the users table, were all sitting there together in a mess that would have been a nightmare to untangle later.

The agent hadn't made a syntax error. It hadn't thrown an exception. It had just built something that looked right, felt right, and was quietly wrong in ways that could have seriously hurt real users. The only reason I caught it was because I understood what I was looking at.

The engineers who will thrive

I've started treating agent output the way I'd treat a pull request from someone smart but inexperienced. I trust the effort, I check the details, and I never assume that because it compiles and runs, it's correct. That mental model has saved me more than once.

The developers who do well in this next period won't necessarily be the fastest coders or the ones who know the most syntax. They'll be the ones with strong fundamentals, good taste, and the ability to hold a mental model of a system while an agent fills in the details. They'll know enough about security to spot what the agent missed. Enough about databases to question a schema that looks fine but isn't. Enough about architecture to know when a technically working solution is quietly building up debt.

The orchestra can play beautifully with the right conductor. But the conductor has to know music. That part doesn't go away just because the musicians got better.