"In two quarters, we're not writing any code anymore." That's what Datadog's co-founder told the whole engineering team, says its CEO, Olivier Pomel. It isn't a ban on code. It flips the job: engineers used to mostly write code and sometimes automate, and now they'll mostly automate and sometimes write.
Datadog isn't an AI lab with unlimited computing power. It's a public company of about 8,000 people that sells observability, the tools that show what's happening inside running software. When a company like that sets a deadline for hand-written code, every team that writes software meets the same question. What is a developer's job when typing the code is no longer the main part?
He said it in a Y Combinator interview published in July. Most of the 30 minutes tells Datadog's story. The part on AI and code takes about seven minutes, and this note is about those seven.
Most companies can't burn a billion
The question about code came right after he talked about AI labs. Datadog counts the ten biggest AI companies among its customers, plus many AI startups and big enterprises like banks and train companies. The top two or three labs use his product "in weird ways", he says. In his words, they have infinite compute, infinite inference and models nobody else has yet.
So the labs don't show what everyone else will do. Most companies, he says, won't be happy to "incinerate a billion in compute" on internal work.
Then the interviewer asked: "Are you token maxing?" Token maxing means spending as many tokens as you can on any task where more work gives a better result. Tokens are the units AI models bill by. I wrote about the word in an earlier note. His answer was short: "We're not token maxing."
Datadog builds the intelligence into its own product instead. Pointing a general coding agent like Claude Code at an ops problem, he says, is "a little bit like trying to send a postcard by buying a business class ticket for it." So Datadog puts that work inside the product, next to the data it watches, so the system can do the work on its own. It's also making the product easy for agents to use, because its customers already point agents at it.
Two ways to spend on AI
The labs can burn a billion, most companies won't
- Infinite inference, in his words
- Models nobody else has yet
- Not what the rest of the world will do
- The intelligence built into the product
- So the system acts on its own
- Easy for customers' agents to use
And still: a deadline for hand-written code.
Mostly automate, sometimes write
The deadline came in the same answer. A couple of months before the interview, his co-founder and CTO, Alexis Lê-Quôc, stood in front of the whole engineering team. In two quarters, he told them, nobody would be writing code anymore.
The interviewer checked: so the rule is no more code? Not quite. "The reality of it is we will write code, we will rewrite code," Olivier says. The point is an inversion: "You used to mostly write, sometimes automate, and now you're going to mostly automate, sometimes write."
He doesn't describe what the automating looks like day to day. My reading: an agent writes most first versions. The engineer decides what to build, checks what comes back and steps in where the agent can't.
He also explains why the deadline is short and the details are loose. The way people write code, "or don't write code", has changed three times in the last year, he says. Nobody could have planned that a year ahead. So the company has to react fast and try many things at once.
The directive, in his words
Writing and automating swap places
BeforeIn two quarters
Mostly write codeMostly automate
Sometimes automateSometimes write code
Not a ban: people still write and rewrite code.
One developer, a couple of days
The deadline came from something people saw, not from a forecast. Like many companies, Datadog had a moment in December. He gives two reasons. The models got good enough: "the threshold was crossed." And it was the holidays, so people had time for a project of their own.
What came out convinced him. "Absolutely excellent developers, like not beginners," rebuilt whole systems alone, in a couple of days. He puts the same work at a team of six for six months, the old way.
That was enough to change how the whole company works. "It's going to be painful. It is painful, but we're doing it."
Rebuilding a whole system
Six people for six months, or one for a couple of days
That convinced the founders to change the whole organisation.
Teams of two or three
The first change he sees is the size of a team. To explore an idea, you used to need a team of eight. Now two or three people can do it.
That fits his rule for this period: AI moves too fast to wait until customers ask for something. So take more shots, and accept being wrong more often. A smaller team makes each miss cheaper.
Everything else is open. "It changes everything," he says, "but the tricky part is we don't know what the destination looks like yet." Nobody knows what the process will be a year from now. Nobody knows how many product managers, designers, security people or engineers it takes to build a product.
So who will the engineers in those small teams be?
People needed to explore an idea
From a team of eight to two or three people
Still unknown: how many PMs, designers, security people and engineers a product needs.
The small tasks were the training
He doesn't talk about juniors. But look at who did the December rebuilds: excellent developers, not beginners. The big gains came from people who already know what a good system looks like.
That knowledge has a history. Juniors used to learn on small tasks: a config change, a first test, a fix in a README. Some of that work went wrong, and that's how judgment grew. Coding agents can now take those tasks.
The hiring data already points that way. A Stanford team follows US payroll data. In jobs exposed to AI, like software development, people aged 22 to 25 stood 19% below where they would be by June 2026. Experienced workers in the same jobs show no such gap. I went through that study in the abundance of intelligence.
One answer is PRAISE, a Microsoft and Carnegie Mellon program. It pairs each new engineer with a senior whose job is to "ship seniors". When a team mostly automates, the small tasks stop teaching judgment by accident. Someone has to teach it on purpose.
How juniors became seniors
Judgment grew from the tasks agents now take
- Small tasksA config change, a first test, a README fix. Agents take these now.
- MistakesSome of that work goes wrong, and you learn why.
- JudgmentYou know what a good system looks like.
The December rebuilds needed that judgment: excellent developers, not beginners.
Try it
The part on AI starts at about 19 minutes into the interview, and the token maxing question comes just before 24.
- Y Combinator, What Top AI Labs Are Really Doing With Observability, with Olivier Pomel, July 2026.
- Spending more tokens only pays with good skills and real tests, on the word token maxing.
- The agentic software factory is the 2022 one with one box swapped, on why review becomes the slow part once building is fast.
- The abundance of intelligence, for the hiring data on young workers.
- A preceptor's job is to ship seniors, on how juniors still become seniors.
Then try his inversion on one week of your own work. Count the hours you spent writing code, and the hours you spent deciding, checking and fixing what a tool wrote. If you try it, tell me which side was bigger.
Plain text: two-quarters-to-stop-writing-code.md · Atom feed · All notes