Adapt or die
If you’re an experienced engineer, writing code is no longer the job. Deciding what to build is.
If done right, AI coding, code reviews, etc. can mostly be done in an automated way. At Google, 75% of all new code is now AI-generated and approved by engineers, up from 50% the fall before. GitHub’s automated code review had over a million developers using it a month into its preview.
Where the good engineer comes in is being able to know what should and should not be built, to build the right guardrails for the right product, and to lead the effort and charge.
A senior engineer right now with a good concept of system design and good, thoughtful architecture can build a lot of the products we see today that took years to make. Teams full of senior engineers ship at an incredible rate: a side project we built AI-first replaced our main product in one month. Therefore, the differentiator is in being able to make a better product, not just a product.
The architect and the general contractor
The engineer is the architect, and the general contractor is the AI.
If you’ve ever tried to get an AI to design a “cool new feature in our current product” without giving it any guidance, it falls flat on its face, at least in my experience.
I once tried exactly that. The prompt was something like this:
Think of 10 new product ideas to build, and send that to a panel of agent judges to decide what would be of the most use to our users. When those are back, rank the top 3, and then do another new set of 10. Do that even another time until you have 9 candidates, have the panel review those, and then build them 3 times each and have a separate review board decide on the final iteration.
It built extra buttons that went to extremely similar versions of pages we already had, a summary page that had no use to the user, and a few shortcuts. This took a lot of tokens and time (I prompted it pretty aggressively to self-review and use agent judges), but it was just a joke. It ended in disaster, with nonsense features that were at best slight variations of what we already had.
A panel of general contractors still can’t do the architect’s job. So, as the architect, your job is to find the things that need building, and have a plan to build them the right way. After that, the AI can orchestrate, spawn sub-agents, do whatever it wants to execute your vision.
The general contractor builds what the architect plans.
The architect before AI
I’ve always thought of myself as a product-minded engineer. There were always (and still are) others around me much better at the craft of “coding”. I think I was able to bridge a gap into leadership by having that perspective, by fighting for better products, not just better code or systems.
Recently, with the advent of AI, that has gone full force. The general contractor at least does it faster and easier. A good engineer would likely build better, more well-thought-out (defensible, patterned, etc.) software, but it would take significantly longer. In one controlled experiment, developers using GitHub Copilot finished a coding task 55.8% faster than those without it. In a randomized trial at Google, AI cut the time developers spent on a task by about 21%.
What the architect actually does
The architect talks to users, kills features, ships and measures. Sure, all of this, but mostly it’s about spending time in your own product, trying to use it, understanding the friction spots and the distractions, and understanding what users want.
This is not out of reach. It just takes some research and thinking.
Too many engineers have spent too long letting product managers do this thinking for them, and that’s been wrong.
Guardrails are the architect’s plans
Guardrails are how the architect’s plan reaches the general contractor. They come in a few kinds: the rules and skills the AI works inside (files like AGENTS.md, which over 60,000 open-source projects use), the architecture and system boundaries, and the tests and automated review that check what comes back.
Sandbox your agents, too. I’ve even run things on machines without any production credentials at all, making sure there’s no possibility for the blast radius to be outside of a crummy PR. Simon Willison calls the dangerous combination the lethal trifecta: an agent with access to private data, exposure to untrusted content, and a way to send data out. OWASP’s guidance points the same way: limit an agent’s permissions to the minimum necessary.
Inside those, the general contractor can do whatever it wants.
Does the architect still need a product manager?
I’ve always preferred hiring engineers who are product-minded, and have thought for a long time that with the right team, a product manager is obsolete. I think that is most definitely true now. Airbnb has already moved that way: in 2023, Brian Chesky said the company got rid of the classic product management function and folded it into product marketing.
The pushback is familiar: an engineer won’t talk to users, won’t own the roadmap, won’t scale, etc. All these are half as powerful as they used to be.
First, just find an engineer willing to get on calls. By the way, lots of product managers I know don’t do this.
Second, teams can be smaller now. Base44, an app builder run by a solo founder with six employees, sold to Wix for about $80 million in 2025. Leadership can own a roadmap as well as a team itself. We can execute so quickly that waiting for someone to “adjust the roadmap” just slows everyone down.
“That doesn’t scale”? That doesn’t scale what? More people? For what reason? Sometimes you’ll need more, but being careful about hiring product-minded engineers will fix this “issue”.
The side project that replaced our main product didn’t have a product manager.
If you hire engineers, hire architects
If you lead an engineering team, the architect is the hire that matters. Give engineers more autonomy, hire product-minded engineers, and get other organizations (such as product) out of the way.
Where to start tomorrow
Becoming the architect starts with three habits.
Use your product daily for a week, and write down every annoyance.
Own one outcome end to end, one metric you answer for.
Automate the craft (agents, rules, reviews) so the general contractor does it and you have time to think.
What happens if you don’t adapt
Not adapting is death: at least death of your career and craft as you know it, in my opinion.
Things might change. Maybe AI will fade because we can’t build data centers, or it will be too costly, but by and large, it’s here to stay.
If you are an amazing coder who knows how to solve very complex problems at scale, some things the general contractor isn’t quite good enough for yet, your time is longer. But I think it’s still shrinking nonetheless. METR found the length of tasks AI agents can complete doubling about every seven months over six years, and faster than that since 2024.
Others have said it many times, but my favorite is Brad Pitt playing Billy Beane in Moneyball: “Adapt or die.”