San Francisco is the new Athens.
I’ve been trying to name what’s actually different about building software right now. The closest thing I keep landing on is ancient Greece.
The GUI era was craft, not philosophy
Pre-AI — call it pre-GPT-3.5 — every product was a GUI. You built screens. Users clicked through them. The interface was the product.
GUIs converge. Tinder invented swipe-to-like, and within two years every dating app had it. Every B2B SaaS dashboard looks like every other B2B SaaS dashboard. The winners weren’t the ones with better design. They were the ones who were loudest. Distribution won. Craft didn’t matter much past a threshold, because anyone with time could replicate anyone else’s good idea.
That era was deterministic. You could look at a product and know how to rebuild it.
The shift to the cockpit
What’s changing is the weight of navigation. People used to drive the software themselves, clicking their way through. Now they sit in the cockpit and talk to it — chat, voice, whatever — while an agent does the actual driving.
You still get the GUI. You still see the dashboard. But the plane is flying itself. The product is moving autonomously, and you’re along for the ride, correcting it when it drifts.
How do you design the thing that’s flying the plane?
Agents are philosophy
When you design an agent, you’re not designing an interface. You’re encoding a belief system about how work should be done. In coding agents specifically, people call this a “harness”. A harness is just a team’s opinion about how the agent should think, what it should see, what it’s allowed to touch, how it recovers from failure.
Look at what’s shipping right now: Factory, Droid, OpenCode, Codex, Cursor, Claude Code. Each one is staring at the same problem — “help people write software with an LLM” — and each one has landed on a different framework. Different tool sets. Different mental models. Different priors about what the user is even trying to do.
None of them are copies. You can’t look at Cursor and rebuild Claude Code. The surface is inspectable; the philosophy isn’t.
Same observation, different beliefs
This is what Athens was. Socrates, Plato, Aristotle, the Stoics, the Epicureans — all staring at the same world. Same problems. What is a good life? What is knowledge? How should we act? They weren’t arguing over facts. They were arguing over frameworks.
You couldn’t replicate Stoicism if you didn’t actually believe it. Memorize every word of Marcus Aurelius, and if your prior was that pleasure is the goal, you’d still end up an Epicurean. The output follows the belief, not the other way around.
Harnesses work the same way. The team’s assumptions about how intelligence works, what humans want, and what software should be leak into every decision. You can’t fake a philosophy you don’t hold.
Customers are chasing ideologies, not features
Here’s what surprised me. Users aren’t picking tools by checklist anymore. They’re picking by which team’s philosophy matches their own.
You meet Claude Code people. You meet Cursor people. You meet Factory people. They talk about these tools the way people used to talk about schools of thought. Not “this one has feature X”, but “this one thinks about the problem the way I think about it”.
And when users can’t find a philosophy they believe in, they build their own. That’s why half of SF is shipping their own personal tooling right now. It’s not NIH. The available ideologies just don’t fit.
Why SF
Athens wasn’t special because of its olive trees. It was special because the philosophers were all in one place, arguing and sharpening each other. You couldn’t be a serious thinker in 400 BC and not be in Athens.
SF is the same. If you’re trying to figure out what software is becoming, you have to be around the people forming new beliefs about it. Reading their essays isn’t enough. The interesting stuff happens in the arguments. (I’ve written about this before.)
Implications
For builders: Stop optimizing for feature parity. Figure out what you actually believe about the problem and ship the most opinionated version of that. Your moat isn’t the feature list — it’s the belief system that generated the feature list. (Related.)
For customers: Pick the tool whose team you’d want to argue with at dinner. In the GUI era, tool choice was utility. Now it’s alignment. You’re buying into a worldview, and your work will start to shape around it.
For investors: Metrics lie longer than they used to. Two teams with identical ARR can have completely different futures because their underlying philosophies will age differently. Evaluate the belief, not just the numbers.
For everyone else: The GUI era rewarded the loudest. The agent era rewards the most opinionated. Figure out what you actually believe and ship it without apology.
That’s the whole game now.
April 20, 2026