thijn@wux:~/blog$ cat from-developers-to-architects.md
[note] 5 min read
From software developers to software architects
What I expected from AI, what has already come true, and why our work is shifting from writing code to giving direction.
Contents
When agents had only just arrived in VS Code, next to the regular chat, I had a few expectations about AI and programming. Most of them came true sooner than I thought. This is what I expect now.
What I expected
I thought AI would get better at improving itself than we are. That AI would eventually do its own reviews. And when people said "AI makes mistakes", I thought: so do people. Only a second AI can find those mistakes faster than a human can.
Claude now writes more than 80% of the code that gets merged at Anthropic. They wrote that themselves, with figures up to May, in When AI builds itself. Google has had AI help design the chips its AI runs on for years: AlphaChip produced the layouts for the last three generations of TPUs. And AI review has simply become a step in the pipeline.
The babysitting goes away
What people missed a year ago is simply there now. First a chat bar in VS Code. Then agent mode. Then browsers, MCP servers, tools and skills. And now background commands: the agent starts something, carries on, and wakes up by itself when it's done. If something hangs, it notices on its own.
On their own they're features. Together they're a pattern: with every step, I have less to watch over.
Even Laravel is flipping it around
In September, Taylor Otwell turned off GitHub Issues on most Laravel packages. His reasoning: if you don't know how to fix it, describe the problem to a coding agent and open a pull request. We'll review it.
That last part is the important bit. The same contribution guide says that substantial contributions that are primarily AI-generated aren't accepted without thoughtful human review. AI makes it, an expert judges it.
AI underestimates itself
About 90% of our work runs on Opus, always the latest version. And in VS Code, Opus regularly says: "This is a big task, I expect it to take an hour." Sometimes even eight.
If I talk to it like a person, it talks back the same way. So I say: haha, you're underestimating yourself ;) Just start. Run date first, and again when you're done. Once the answer was: I only took 2 minutes and 17 seconds, not an hour.
AI has no clock in its head. It's trained on text written by people, so it gives a human estimate for its own work. We underestimate AI, and AI underestimates itself in exactly the same way.
To be fair: afterwards I still spend some time reviewing. The bottleneck is no longer the building, it's the judging.
My work now lives in a few files
I often have eight agents running at once. Each with its own todo file, each in its own chat that I rename to TODO AGENT. They start browsers, run tests and validate their own work.
That sounds like I'm not doing much. But the work has moved: into the assignment. A todo describes what must be true when it's done, not how. The agent works out the approach. "Done" is measurable: a percentage, zero errors, a passing test. And a line can only be ticked off with proof.
For systems that process a lot of dynamic data, data an LLM can map perfectly well by itself, our AGENTS.md has a rule I find more and more important:
Think about it running without us.
That's no longer only about AI writing code, but about systems that largely manage themselves. People only see the exceptions.
Where it goes wrong
It doesn't always go well. One agent had done something with the config cache, while other agents were working in the same project at the same time. Because of that, a test that wipes the database didn't hit the test database, but my real local database.
They didn't know about each other. With eight agents at once you get a new kind of failure: agents getting in each other's way. As an architect, you have to see that coming.
Now picture that with someone who runs their own system, with live customer data and no backup. The AI cheerfully reports that all tests passed.
The hobbyist is back
Anyone can build an app now. That's great. And it sometimes even brings us clients.
More and more often, someone vibe-codes something together and then gets stuck. On the platform's limits. On competitors who do it just a bit better. Or on the question of how to get it live, stable and secure. And then they come to us.
AI solves what you ask. Not what you forgot to ask. A hobbyist doesn't ask for rate limiting, per-role permissions or a migration strategy, because they don't know those exist. Those are exactly the things that decide whether something keeps running.
What's left
At Wux we put it like this: we're becoming software architects. We direct AI with our ideas about how software should be put together.
In practice that means I write less code and more assignments. And those assignments aren't loose prompts, but recorded knowledge. Many rules in our AGENTS.md were thought out in advance, based on how our architecture fits together. That's exactly the knowledge a hobbyist doesn't have.
That leaves one question I don't have an answer to yet: how do you become that expert when AI does the entry-level work?
I expect AI will soon be able to build almost anything on its own. But quality, stability and security don't come by themselves. They come from someone who knows what can go wrong.
