A talented engineer with agents now ships more than a hundred of them.

Is the rockstar back?

On 11 July 2019, Shekhar Kirani, a partner at Accel India, posted an eleven-tweet thread telling founders to grab a 10x engineer if they ever found one, then listing how to spot them. The traits were a caricature: hates meetings, keeps highly irregular hours and codes late at night, knows every line in production, black laptop background, worn-out keys. The reaction was fast and hostile. The dominant read was that he was describing the brilliant jerks the rest of the industry spends its careers cleaning up after, and TechCrunch described the scorn as tsunami-like. It became a meme within days.

What actually happened, beyond the dunking, is that the thread killed the term's respectability. The argument had been simmering since the early 2010s, when job ads asked for "rockstar ninja developers" and got mocked for it, but Kirani gave the sceptics a single concrete document to point at. After July 2019 you could not use "10x" or "rockstar" sincerely in serious company without a footnote. The industry's consensus settled roughly on: the productivity spread is real, the personality attached to it is a liability, and if you're hiring for the personality you're doing it wrong.

Is the rockstar back? I hope not. The traits that got mocked deserved it. Refusing to document, refusing review, treating every colleague as an interruption: those were never the source of the output, they were the cost of it, and the rest of the team paid. Give that person a fleet of agents that never argue and you don't get a 10x engineer, you get bad decisions shipping at machine speed with nobody in the way. Leave the rockstar dead.

Is the surgeon finally getting its spot in the sun?

In 1975 Fred Brooks described a team organised like an operating theatre, borrowing the idea from Harlan Mills, who had proposed the chief programmer team at IBM a few years earlier. One surgeon does all the cutting: specifies, designs, codes, tests, documents. Nine people exist to make sure the surgeon never puts the scalpel down. A copilot to argue with the design. A toolsmith, a tester, a program clerk keeping the code as team property, someone to handle the organisation, someone to make it all legible. One mind on the problem, many hands around it. Brooks was clear this produced better software than a large team each hacking at their own piece. He was also clear that almost nobody did it.

The reason was never that it was wrong. At the same payroll, ten average engineers on ten features shipped ten things, and a surgical team shipped one person's worth of design, done properly. Nine of those ten salaries produced nothing a customer or a finance director could point at. Surgeons were scarce, so you couldn't build fifty teams even if you wanted to, and when your one surgeon left the team was dead. Volume was the safer buy at the same price. So the industry bought volume, put bums in seats, and accepted that the software would be worse because there would be more of it, sooner.

That is the trade AI broke. The copilot is now called Copilot. The language lawyer is a model. The toolsmith, the tester and the clerk are agents that never tire, never leave and cost a few hundred pounds a month. Nine salaries collapsed into a subscription, and the surgical team's output per pound flipped from the worst deal in the building to the best. The surgeon didn't get better. The nine people around them got free.

So is the surgeon finally getting its spot in the sun? Logic says that it should. The trade that kept it on the bench for fifty years was a cost argument, and the cost is going: some of it gone already, the rest on its way as models and the harnesses around them commoditise. What's left to argue about isn't the price. It's whether anyone remembers what the other nine were for.

What were the other nine for?

Cheap, then. One seat, a subscription, and the surgeon gets everything Brooks said they needed. Except that the nine did four things, and the subscription replaced one of them.

Hands. Writing the tests, building the tools, keeping the records, doing the grind that kept the surgeon on the design. That's what agents do, and they do it better than the nine ever did: no fatigue, no notice period, no opinion about the ticketing system. This is the part everyone has noticed, and it's the part that makes the surgical team look free.

Friction. Brooks' copilot was there to argue with the design. His tester was described as an adversary. Nine people meant nine chances for someone to say "why?" before the thing shipped, and the surgeon had to sit through it because those people had salaries and opinions. Agents will argue if you ask them to, as hard as you tell them to, and you can dismiss them with a keystroke. Friction used to be structural. Now it's opt-in, which means it's available precisely to the people who least need it.

Memory. Nine people who knew why the system was shaped the way it was, which decisions were deliberate and which were Tuesday. That knowledge now lives in one head, or in whatever the surgeon bothered to write down. When the surgeon leaves, both go.

Pipeline. The copilot was explicitly the surgeon minus ten years. That seat wasn't just support; it was where the next surgeon came from. Delete it and the surgical team has no succession plan, and the company that runs ten of them has no idea where the eleventh will come from.

So the company that goes small has bought a quarter of the surgical team and needs to know it. If it remembers what the nine were for, it keeps one person with no stake in the design who reads it anyway, it writes things down as if the surgeon might leave, and it grows the next surgeon somewhere on purpose. If it doesn't, it thinks it has a surgical team and actually has a rockstar with a subscription.

So who wins now?

Here's what I think happens next. Headcount stops being a proxy for output, and then stops being a proxy for anything. The teams that ship the most will be small, senior, expensive per head and cheap per result. The seats that go first are the well-paid ones in the middle: enough skill to be trusted with a feature, not enough judgement to be trusted with the design. The floor of the profession doesn't get lifted, it gets removed. And the companies that keep counting seats will keep looking healthy right up until the quarter when a team a fraction of their size ships their roadmap.

The bar for a bum to get a seat has risen, and it will keep rising for a while. The number of seats it takes to support a product or an experiment has collapsed. And there are whole tiers of skill and whole types of personality, the second-class rockstar chief among them (the traits without the output), whose value has dropped through the floor: knowing an API by heart, typing without pausing, being the only person who could make the thing go. Those were scarce. They are now a subscription.

I'm not claiming to be first here; I suspect I'm not even early. A lot of leaders reached this conclusion years ago and threw their dice. However disliked some of them are, and however little skill or empathy they showed in making the cuts, it's hard to argue they aren't winning at business.

There's an obvious objection. Capability at lower cost means we can afford experiments that were previously too risky or too small to staff, so the seats don't disappear, they move to new work. That's fair, and it will happen. But there's only so much money in experiments that used to be unaffordable without agents. The cheap end of the market got bigger; it didn't get rich. The seats that go were supporting expensive bets, and no cheap experiment replaces those.

The reality

That's the economics. The people side will be messier, and, as always, it will follow whatever brings the most money in the easiest way imaginable. Or whatever people think will.

The rockstar will stick around. It was always the easiest way to bootstrap a startup; the problem was that at some point their personality became more of a liability than their output was worth, usually around the time the company needed to scale. With agents, that point moves. The faster typewriter buys them a longer runway, and the reckoning arrives later, and larger.

We will see more surgeons. Now that the hands are affordable, the maths that kept them rare has flipped, and I expect the number to follow the maths.

Corners will still be cut, and many of the people we call surgeons will be rockstars in disguise. Sometimes because of their behaviour. Far more often because of the organisation around them: nobody funded the reviewer, nobody insisted anything be written down, nobody was growing the next one. The surgeon didn't fail. The company bought a quarter of the team and called it done.

What happens to the people arranging the seats?

As delivery teams get smaller, they need less orchestration and fewer layers of management. There's no version of this where the management layer stays the same size while the teams under it shrink.

So managers will have to pick, according to their abilities and their preferences. The bar will rise for them as well; that's not in question. Do you still enjoy the work, and are you actually good at it? Then keep doing it. You'll prompt your observability agent to work out what's going on in prod instead of asking the on-call engineer, and you'll spend more of your week on the three things nobody else is funding: the reviewer, the writing down, the next surgeon. That is no longer the boring half of the job. The manager who funds the friction is the only thing standing between a company and a hundred rockstars with subscriptions.

Or will you risk it and try to be a surgeon? It's a good time to, but good luck not turning into a rockstar; "you're absolutely right" plays so sweetly on the mind, and the agents say it every time. Gide had it right seventy-odd years ago: believe those who seek the truth, doubt those who find it. The agents have always found it. Always start a new session when you hear those sweet words, and remember what the other nine were for.

One thing I am certain of. If you're still playing the headcount game, proudly counting the seats, you're in for a disappointment. That asset is turning into a liability, fast.

AI usage disclaimer

English is not my first language, and I'm not a talented writer in any language, so I use AI to polish my writing. The ideas, the setup, the mistakes and the opinions are all mine. They reach the AI as thorough notes in half-broken English, and it fixes the English. It doesn't supply the thinking, though it is a trusty research assistant.