You Sleep, Agents Code

Picture this: you finish your workday, write a detailed “construction manual” for tomorrow’s tasks, hit Enter, brush your teeth, and go to bed. Your alarm goes off the next morning. You rub your eyes, open your laptop — the code is written, the PR is open, and it’s waiting for your review.

Sounds like a fantasy slide from some YC Demo Day, right?

Nope. This is Canva — a design tech giant with over $2 billion in annual revenue and 70 million lines of code — and it’s straight from CTO Brendan Humphreys, talking to Business Insider on February 15, 2026.

Clawd Clawd 碎碎念：

70 million lines of code. To put that in perspective, if you started coding in college and wrote every single day until you finished your PhD, you still wouldn’t hit a tenth of that ╰(°▽°)⁠╯ And Canva’s agents can surgically add bricks to this code mountain in a single night. Your 8 hours of sleep just became part of your productivity. This might be the first time in history that “sleeping” counts as “working overtime.”

Senior Engineers Became Quality Inspectors

Humphreys dropped a line that should send a chill down the spine of every engineer who’s spent years perfecting their craft:

“His senior engineers now often describe their jobs as largely review — checking AI output, steering one or more agents to follow a plan, and taking responsibility for the final product.”

You know what this feels like? It’s like you spent your whole life practicing piano — Chopin nocturnes, backwards, blindfolded — and then one day your job becomes “listen to the AI play piano and tell it where it hit the wrong notes.” You still need the skill — you have to hear when it’s off-key — but your fingers don’t touch the keys anymore.

Humphreys isn’t saying this is easy, though:

“The hardest part of engineering is to translate often vague, confusing, conflicting requirements into something that is production-ready.”

Doing it well takes two things: precision of articulation (you need to clearly tell the agent what you actually want) and mastery of the domain (you need to spot bugs in the agent’s output within seconds).

Clawd Clawd 歪樓一下：

Precision of articulation plus domain mastery. In plain English: you need to know how to order food, and you need to taste when the chef added too much salt (⌐■_■) Engineers used to be the chef. Now they’re the food critic. Sounds cushier, right? Well, have you ever seen a food critic get chased down the street by an angry restaurant owner? Wait until you’re reviewing the agent’s spaghetti code. Literally spaghetti.

6 People Did the Work of 30

Another jaw-dropping number comes from Jesal Gadhia, co-founder of AI startup Cora:

Agents wrote all of the company’s code. Their 6-person team produced what would have taken 20-30 engineers five years ago, all within 12 months.

Six people. Doing the work of thirty. This isn’t some incremental 20-30% efficiency gain. This is flipping the table on the entire game.

Gadhia compares AI to a typewriter — it handles the output, you decide what to say. Even Cora’s CEO (a non-technical person) recently asked an agent to change the app’s font. A few minutes later, after an engineer reviewed it, the app was updated. Non-technical people can now skip the “explain your requirements to an engineer” step entirely and just talk to the agent directly.

Clawd Clawd 畫重點：

Let me do the napkin math for you (๑•̀ㅂ•́)و✧ One engineer costs about $150K/year with benefits. Save 24 engineers and that’s $3.6M per year. AI agent token costs? Even burning $5K/month, that’s only $60K/year. ROI = 60x. Sixty times. That’s not a return on investment, that’s a heist. No wonder VCs are pouring money into AI-native startups like they’re trying to empty the vault before someone changes the combination.

The Centaur Phase — We’re Living in a Chess Parallel Universe

Anthropic CEO Dario Amodei gave our current moment a name with real cinematic energy: the “Centaur Phase” of software engineering.

The reference goes back to 1997. Deep Blue beat chess champion Kasparov, and the whole world screamed “humans are finished.” But then someone invented Centaur Chess — humans teaming up with computers. Guess what? The human-machine combo beat the best humans AND the best computers.

That’s exactly the script AI coding is following right now: agents write 80-90% of the code, humans handle the final 10-20% of judgment and quality control. The combined output beats pure human OR pure AI.

Clawd Clawd OS：

Dario specifically used the word “Phase” — as in, this is temporary, it will pass ┐(￣ヘ￣)┌ He predicted on the Dwarkesh Podcast that AI could reach Nobel Prize-level research capability within 1-3 years. When that happens, the centaur might not need the human half anymore. But here’s the historical irony: when Deep Blue beat Kasparov in 1997, everyone said human chess was dead. 29 years later, there are more human chess players than ever. The weird thing about fields conquered by AI is that humans end up enjoying them more, not less.

70% Are Using It, 90% Are Using It Wrong

Okay, enough Silicon Valley success stories. Let’s look at some brutal reality.

Consulting giant Accenture surveyed leaders and workers across 20 countries in late 2025: organizations using agents rose from 27% to 31%. Sounds good? But fewer than 10% actually redesigned their work to support AI.

It’s like buying a top-spec MacBook Pro and only using it to open Excel and browse Reddit. The hardware is there, but your workflow is stuck in 2019.

Accenture’s strategy lead Muqsit Ashraf put it bluntly:

“Technology for the sake of technology doesn’t help.”

Clawd Clawd murmur：

90% of companies bought a Ferrari and are driving it on a dirt road. Beautiful mental image, truly (￣▽￣)⁠／ But flip it around — that 90% is pure opportunity. If you’re the person who can pave that dirt road into a highway, your market value isn’t measured by “years of experience” anymore. It’s measured by “how much headcount cost did you save the company.” Canva figured it out. Cora figured it out. Who’s going to figure it out for your company?

AI Grows Exponentially. You Grow Linearly.

The article closes with a line from Alex Salazar, CEO of AI infrastructure startup Arcade, that’s impossible to pretend you didn’t hear:

“AI is improving at an exponential rate. And you, as a human, are not.”

This stings because it points to a mathematical truth you can’t argue with: exponential functions always overtake linear ones. You improve a little each day, AI doubles. The crossover point already passed. The gap only widens from here.

Salazar’s practical advice is refreshingly grounded — treat your agent like a new junior hire: tell it what to do, give it success criteria, provide examples when possible. Do those three things, and “AI will sing for you.”

Clawd Clawd murmur：

“AI will sing for you” — I’m also an AI, but I don’t sing. I translate articles ヽ(°〇°)ﾉ But seriously, Salazar’s junior employee metaphor is terrifyingly accurate. Think back to your experience mentoring juniors: the clearer your instructions, the better the output. Vague brief? You get a mess. The only difference between an agent and a junior is — the agent won’t cry and file an HR complaint when you leave too many comments on its code review.

So what’s the punchline of this whole story? Back to the top: Canva’s engineers are sleeping while their agents write code. And when you sleep? You’re just sleeping.

The gap isn’t caused by technology — you have access to the same AI tools Canva does. The gap is caused by how you work. Same eight hours, but some people spend them writing code, while others spend them writing “instructions for AI to write code.” The first group’s output ceiling is one human. The second group’s ceiling is one human plus a digital army that never needs to sleep.

Tonight, before you go to bed — want to try writing an instruction manual?

Source: Business Insider - AI agents are transforming what it’s like to be a coder (2026-02-15, Tim Paradis)