Sixteen years in, AI writes my code. Here’s what I’m still for.

Posted on Oct 25 • Originally published at dineshkumars.dev

I wrote this with AI assistance. That’s part of the point: the sources, experiences and final calls are mine, which is exactly the human role this post is about.

Uber burned through its entire 2026 AI coding budget in four months. Not one of its riders ever asked for AI. They asked for a ride that shows up.

That’s the gap this post is about. Everyone is paying for AI now, per engineer, per seat, per agent left running overnight. But customers never buy AI. They buy something that works, and someone to call when it doesn’t.

I’ve felt the shift from the inside. I’ve been a software engineer for 16 years, and for most of them my value was measured by what I could build with my own hands: code, systems, designs. Today I work on an agentic AI platform at Trimble, where AI agents run and execute real work. The first time AI did a day’s work for me in minutes, I was mostly excited. Now most new code around me is written by AI, the bills keep growing, and the question I keep asking is what I’m actually for.

Working on that platform has taught me one blunt lesson. An agent with no business value is not something any customer will pay for, however clever it is. It just becomes a void: tokens in, nothing out.

Four talks helped me answer it, from a mathematician, an npm co-founder, the head of Google Research and an engineer who builds agent harnesses, along with this year’s news on AI costs. They don’t agree on everything, but they point the same way.

My short answer: AI doesn’t make engineers obsolete. It moves our value, out of typing code and into deciding what’s worth building and paying for, proving it’s right, and growing the people around us. This field moves fast, so treat the post as a snapshot. More on that later.

The first thing that helped me make sense of my own job came from an unexpected place: a math video. Grant Sanderson of 3Blue1Brown made There’s never been a better time to go into math for math students afraid of AI. It turned out to be the best lens I’ve found for engineering too.

Grant maps math in three columns: what mathematicians do, why math is valuable, and who pays for it. The surprise is the last column. Most of the money comes from teaching, not from theorems. Our version looks like this.

None of this is new. Customers never paid for code, even before AI. They paid for a problem solved, and for someone accountable when it wasn’t. Code is to us what proofs are to mathematicians: the raw material, not the product. What AI changes is how easy it is to forget that. When code arrives this fast, it’s tempting to treat more code as more value.

Everything else in this post follows from that. Code was never the product, so producing it faster doesn’t automatically produce more value. It just produces more code, and more code has to be reviewed, run and paid for.

Before we go further, a small game. AI is moving so fast that some of what sounds laughable today may be ordinary in a few years. So imagine a parallel world called Earth 2201, where the wildest possibilities have already happened. We’re not there. We’re on the timeline ship heading toward it, and nobody knows how long the trip takes.

Why 2201? No reason, except that it’s far enough away to be funny and close enough to be a little uncomfortable.

Every so often in this post, you’ll get a postcard from there, marked “Meanwhile, on Earth 2201.” Read them lightly. The serious argument lives everywhere else.

When AI started writing most of the code around me, the first thing I noticed wasn’t the speed. It was the review queue. Laurie Voss, who co-founded npm, put numbers on that feeling in his talk AI Writes the Code. So What’s Left for Engineers?

Teams using coding agents produce 741% more code but ship only about 30% faster. The gap is review. A careful person can review roughly 450 lines an hour, and no model has changed that. AI reviewers help, but different models flag different bugs on the same code, and none has what Laurie calls the maintainer’s instinct: the history and side effects that live in someone’s head.

This is exactly what I see. The biggest problem AI code causes around me is that there’s too much of it to review properly. We’ve had a close call where AI output was confidently wrong, and a person caught it in review. The safety net is a human, and that human is already overloaded.

Laurie’s bigger point is the theory of constraints. Fix one bottleneck and the next appears: review, then operations, then security. The one he thinks agents won’t eat is deciding what to build and what “good” means. That isn’t in any training data. It’s in customers’ heads, and someone has to go and talk to them. He calls that person a product engineer.

Meanwhile, on Earth 2201. I can imagine a world where the customer is itself an AI agent, thinking through what it needs and handing requirements straight to product agents, with no humans in the loop at all. It’s far down the line, not today, so no need to worry about it yet.

His practical advice is to spend human attention where mistakes are expensive. This is how I’d sketch it.

This talk was the closest to my day job, because an agentic platform is, in large part, a harness. Jediah Katz spent years building the agent harness at Cursor, the layer that wraps a model and connects it to real tools. His talk, Smarter Models Killed the Harness. Here’s What’s Left, explains why so many clever tricks stopped mattering.

A harness adds guardrails so a model can try harder tasks. People use it, that creates data, labs train on it, and the next model just knows how. Then the guardrail gets deleted. His team even has a bot that opens a pull request to remove a hack once it stops firing. On frontier models, he says, a fancy harness and a bare-bones one now score about the same on coding benchmarks.

But some things no model absorbs, however smart it gets.

An agent that runs for days needs infrastructure. Security teams won’t hand a model every key and hope for the best. And the harness is where cost gets controlled: caching and mixing models can make the same work much cheaper to run. So cost per task is now an engineering problem, and a whole industry is forming around it. It’s Laurie’s bottleneck in a different costume. The clever part gets absorbed, and what remains is trust, judgment and cost.

Meanwhile, on Earth 2201. AI has quietly abstracted programming languages away, the way compilers once hid machine code behind languages like C and Python. Agents now write in forms no human ever learned, and humans can’t build software without an agent anymore. (To be fair, some of us on Earth 2026 already feel that way.) There’s only one way to build things: describe what you want, and check what comes back. Which, if you think about it, is the whole job this post describes.

On a platform where agents run all day, cost isn’t an abstraction for me. Jediah’s point about cost deserves a section of its own, because this year it became impossible to miss. Per-token prices keep falling, and code may well get cheaper in time. Right now, though, agents use so many more tokens that the bills are going the other way.

Uber’s four-month burn wasn’t a one-off, and its COO has said it’s hard to connect that spend to more useful features for customers. Microsoft revoked its developers’ Claude Code licenses months after enabling them. In TechCrunch’s reporting, the head of the FinOps Foundation describes companies finding themselves 3x over their whole 2026 token budget by April. A Priceline employee says a Cursor renewal came back 4 to 5 times more expensive, and Jellyfish found per-developer consumption rose about 18.6x in nine months. One CTO said an engineer spent $40,000 on tokens in a single month, and he didn’t know whether to stop him or tell everyone else to copy him.

The result is a new line on the cost of every person who uses AI at work. Start with engineers.

Two engineers on the same salary can now cost very different amounts, depending on which models they reach for, how many agents they leave running, and how often they regenerate instead of thinking first. Spend has become a judgment call each of us makes every day, and sooner or later someone asks whether it paid for itself.

And it isn’t only engineers. Sales, product, operations, support, project managers, VPs and the CEO all use AI now. Uber’s CEO has described uptake across its legal and marketing teams as well as its developers, and Uber has since put all 30,000 employees on ChatGPT Work. Every one of those seats adds its own variable bill.

So the company’s AI bill now grows with the whole headcount, not just the engineering team. And outside engineering, there’s often nobody whose job is to ask whether each use is worth what it costs.

That makes “can we build it?” the wrong question. The answer is almost always yes. The better questions are what to build, what not to build, what to redo and what to leave alone, measured against the revenue or value it brings today.

Rewrites need the most discipline. When an agent can regenerate a whole service over a weekend, it’s tempting to redo everything. But rewriting something that already works and earns money burns tokens and review hours just to get back to where you were, and adds new risk. I’d rather redo only what’s costing us now, in incidents, slowness or lost customers.

Meanwhile, on Earth 2201. Every engineer has a token allowance like a phone plan, and unused tokens roll over to next month. The most respected engineers are the ones who shipped the most with the least, and some teams proudly hold zero-token Fridays.

There’s a harder side to all this, and it isn’t hypothetical. When every role carries a salary plus an AI bill, companies start asking a blunt question: is this person adding value an agent can’t? Where the answer looks like no, some are already acting on it.

CBS News has tracked the companies tying layoffs to AI. Block announced it would shrink from about 10,000 people to 6,000, with Jack Dorsey arguing that a much smaller team using its own AI tools can do more. Coinbase cut about 14% of its workforce as it moved to agents and consolidated jobs. Dow is cutting around 4,500 roles as it steps up AI and automation. And by the count of outplacement firm Challenger, Gray and Christmas, US companies directly cited AI in 55,000 job cuts in 2025, more than 12 times the number two years earlier.

Not all of it is what it looks like. Some economists in the same reporting suspect companies are using AI as a convenient story for cuts they wanted anyway, and that its bigger effect so far is on hiring rather than layoffs. But for any one person, that’s little comfort.

So I take it seriously. The safest place to stand is doing work an agent can’t do yet: deciding what’s worth building, proving it’s right, and earning the trust of the people around you. That’s not a slogan anymore. It’s the line companies are drawing.

Meanwhile, on Earth 2201. Every annual review asks one question: what did you decide this year that an agent couldn’t have? Pay is set by the answer, and nobody lists lines of code anymore.

I keep asking myself how far this goes. Grant pushes the question to its extreme: what if AI keeps taking more?

If AI writes the code, the work becomes framing problems and checking answers. If it designs the systems too, someone still has to understand the customer’s world and live with the trade-offs. Grant’s gym analogy fits here: a forklift lifts more than you do, and people still go to the gym.

And if AI does all of it, Grant says teaching survives, because people learn from people they respect. Our version is accountability. When something breaks, a customer wants a human name, not a model version. With each step the human part gets smaller. I think it also gets heavier.

Meanwhile, on Earth 2201. Writing code by hand has gone the way of the abacus. Calculators never killed the abacus; kids in many countries still go to abacus classes to sharpen their minds, and the best compete in mental-math championships. On Earth 2201, children go to weekend hand-coding classes for the same reason. Nobody ships software that way, but the fastest kids compete in tournaments where agents aren’t allowed. Grant’s gym, in its final form.

After the cost and layoff numbers, the most hopeful voice I found was Yossi Matias, who heads Google Research. In this interview he talks about the “organization of one”: each of us carrying a small company in our head, with AI as a virtual lab doing the heavy lifting. The scientist becomes something like a lab director. Swap “scientist” for “engineer” and it’s Laurie’s product engineer, almost word for word.

He’s also clear about which skills count now.

He makes a point about trust that I keep coming back to. Google Translate went from a novelty to something we just assume works, once it was trustworthy enough. The best technology becomes invisible. And his biggest worry isn’t AI itself. It’s choosing the right problems when you can solve almost anything, which, given the bills above, is also a question of what you can afford to solve.

This is where the talks disagree, and where my own view is strongest.

Laurie worries we’ve stopped teaching. Juniors used to learn on small, well-specified tickets, and those now go to agents. He cites data showing employment for engineers aged 22 to 25 falling while demand for engineers in their forties rises. Yossi is more hopeful. With vibe coding, people without a CS degree can build real products, so maybe the ladder has just moved.

Around me, I see both. Fewer juniors are being hired. But the ones who get in are thriving with AI, learning faster and taking on bigger work earlier. So I think Laurie is too pessimistic. The door is narrower, but there’s more room once you’re through it. Someone still has to teach the new first rung, though: asking a good question, writing a clear spec, checking what the machine hands back. Colleges won’t get there quickly. A lot of it falls on senior engineers.

Meanwhile, on Earth 2201. Juniors apprentice to agents for their first year, and the agents write their performance reviews. The reviews are said to be very fair, and a little too honest.

Outside the US, the risk to jobs can feel bigger than these mostly Silicon Valley talks suggest. In many places, a lot of engineering work has been well-specified delivery work, which is exactly what agents are getting good at. For engineers anywhere whose work looks like that, moving up the chain isn’t optional anymore.

Watch AI talks from 2023 and you’ll see how fast this moves. Even the job title keeps changing. Jediah traces it from prompt engineer in 2023, to context engineer in 2025, to harness engineer in early 2026.

Cost will move too. As models and harnesses get more efficient, code may well become genuinely cheap, and some of this post will need rewriting. Beyond that, once AI is woven into the whole economy, the disruption will be global, and I suspect the questions will change, not just the answers. Jediah predicts that 90% of complex agent work will run in the cloud within six months. Read this as a snapshot from October 2026.

Meanwhile, on Earth 2201. Down this timeline, the job title has changed a dozen more times, code really is cheap, and someone is reading this post the way we read old predictions about the paperless office. We’re just on the ship getting there.

Meanwhile, on Earth 2201. The value test from earlier has been automated. An agent now decides what each human is worth to the company and, quietly, to humanity. The humans who still matter are the ones who argued with it, and won.

If Earth 2201 feels far, the timeline keeps going. Science fiction has been sending us postcards about humans and AI colliding for decades. A few for the road:

Every one of these is a branch, not a destiny. Which one we drift toward depends on the very things this post is about: who decides what’s worth building, who checks that it’s right, and whether people can trust the result.

So which branch? I can’t choose for the world, only for myself. Putting everything above together, I’m betting on three jobs that still need us, and these are the seven moves I’m making toward them.

I’m not starting from zero. I already speak and present more, mentor juniors, write papers and design docs, and spend time with customers. None of it is finished, and all of it needs more of me.

A few moves deserve a word. “Own the domain” now includes knowing what’s worth paying for, and what not to build or rewrite just because we can. “Define what correct means” is really trust and safety for AI systems: can you show, not just claim, that the thing behaves? And “keep building by hand” sounds old-fashioned, but it’s how I can tell when the machine is confidently wrong. It’s also why I’m back to studying the math behind generative AI.

Grant’s answer to the PhD student was simple: take a long walk with a friend. I think we should actually do it. Ask what you enjoy most about your work, then what you want to be known for three years from now. Come home and write a paragraph on the second.

That paragraph will probably pull you toward architecture, the domain, trust and safety, or growing people. Any of them will hold up. Pick the one that gives you energy, and, as Laurie says, measure yourself by what ships and gets used, not by how much code you produced or how many tokens you burned.

What would your paragraph say? I’d genuinely like to know.

This post is a small example of its own argument. AI helped me draft it, shape its structure and build the pictures. But a human chose the four talks, pushed for sources on what AI coding really costs, answered the hard questions from real experience, rejected framings that weren’t true, and decided what stayed in.

That’s how most products will be built now. The machine does much of the making. The judgment and the trust are what the human brings, and they’re built into the result.

Written with AI assistance. The sources, experiences, opinions and final calls are mine.