It's sort of like a sickness — once you get exposed to it, you can't stop. I did my first startup back in 1996. I know, I know, it's old. But the whole idea is you come up with a premise and you want to see it come to life. That first one, I had to raise about $11 million. It took a whole group of people to get it done. Every one after that took less people, less money, and less time. Now we're at this huge jump — we're moving into agentic systems where they're basically writing the code. The current constructs I'm working with are writing almost 100% of it. I'm programming right now in Python and Next.js, and I really don't know them — but I've been a CTO, I know how to get teams, I know the product I want, I know the guardrails. I just don't know the languages anymore.

I've never seen a break like this, ever. What it causes you to do is go through a lot of the ideas you never went through — challenge them, work on the core idea, see if it's valid before you commit resources. For me it's the most powerful, most creative tool I've ever had. I've run organizations with 125 people reporting to me. If I had an idea, it would take months to get through all the layers to execute it. Now I can execute it in weeks, if not days. I think the big question now is: how good are your ideas?

Typically in every startup I've done, you hire archetypes — a build engineer, a front-end engineer, a full-stack engineer. Now you don't need to; you can create a hybrid of multiple skill sets in the same agent. But it still comes down to the same best practices we've always used. When I do development, I give guardrails to all my agents. Each one has a profile — a markdown document that tells it who it is and what it should do — and then I build gates around all of that. I use Claude, hooks, and other mechanisms to create those gates, because the agents will still hallucinate. They'll say "no, no, I'm done, it's good" — and it's not done, it's not good. But it also lets me finally do test-driven development properly, which I've always wanted but could never get a team to really commit to. Now all my agents do TDD, and I have a gate that checks the tests exist before any code runs. Most of my time now goes into translating what I'd call the best way to develop software into gates and guardrails the agents can work inside of. You also can't be too rigid — you have to give them room to breathe, but keep a feedback loop telling them if it's working.

If you've been a CTO, or you understand how to develop large-scale enterprise software, these tools are really good for you. It's hard for someone who hasn't felt the pain and the bruises to know exactly how to do it — so it's challenging for people just coming into engineering. That said, we're taking people fresh out of school and training them in this approach. From my perspective, you've got to drink the Kool-Aid. If you fight it — if you say "I'm going to check every line of code" — it just doesn't work, because the speed is too fast. If you think you have to review and understand everything, you don't get any power out of it.

It's a brilliant question. We have one engineer in his late 20s who gets it — he moves fast, creates incredible stuff, very focused and driven. And I've got others who are younger who don't get it, who just move around without landing anywhere. And I've got older engineers where I think it really is a mindset. Age doesn't matter, but you need the right mindset, and you need enough experience to know things like test-driven development — otherwise you won't get what you want out of it.

There are two dimensions to this. One is the models. Up until Claude Opus 4.5, I was trying everything — Llama, open-source models — and they just couldn't stay focused. Not coherent long enough, context window too short. It would go off the rails, forget what I told it, and go do something else — like a monkey running around with a drill. I had to keep it boxed in tightly. With 4.5, that changed. It stayed coherent long enough to actually produce quality code. I still had to shrink the context — everything had to fit inside the 200K window. Then with 4.6 and 4.7, we got to a million-token context, and that changed everything again — now there was enough context for it to just work.

The other dimension was me — I wasn't structuring my ideas correctly, and I didn't have enough guardrails in place. Some areas need to be very structured and directive, others need room to be loose. Once I got that right, it took off. I could have agents work for a half hour, an hour straight — I could run ten to twenty agents at once, all working. It was a little frightening, honestly, but it worked. If you view AI as a line that keeps climbing toward an asymptote, then the problems I had before 4.5 should just keep disappearing — and that opens the door of opportunity way up.

Huge ripple effect. You already see Salesforce saying they're moving to be an API company — getting out of the UI game. In the old world, we took generic products, put them together, and wrote huge integration layers to get them to talk to each other — an endless focus. Now, as these systems go agentic with an MCP interface, a CLI, or an API, you just string them together. They don't dictate the UI at all — you build whatever UI you want, even your own Slack-like tool if you want one. It's not "I have one product for everybody" anymore — I have a product that fits you. And the economics are radically different. A team of ten quality engineers in the Valley, annualized, is millions of dollars. Do it through something like Claude Code — a couple hundred bucks a month, a few max plans cycled through — and you're talking peanuts. It won't be quite as good yet, but the code keeps getting better and better. I think the world goes very bespoke, and it's very hard for SaaS to compete — they're selling this buffet that's all tied together and can't flex. Their cost structure is going to be enormously problematic when you've got thousands of employees to support.

I'd almost disagree. What I see now is that if you build on stacks using something like shadcn and Tailwind, you get better design than most people could do on their own — all you're doing is customizing it. There's a product called Pencil, for instance — designers build the UI, then export it straight to React with shadcn baked in. That used to mean designing, exporting assets, writing manifests, all of that. Now it's just done. The building blocks are getting more Lego-like, more snap-together. What used to be a bunch of separate steps keeps collapsing into fewer steps — and every time it collapses, it gets faster and better.

Honestly, I don't quite know yet — it's still too early to tell. But we look for talent anywhere. We have an offshore team in Serbia, highly talented people — one of the engineers we talked about earlier is on that team. We need to find great people wherever they are.

We're online at everybite.com. We're on some of the biggest restaurant chains — Steak 'n Shake, for one — powering that experience for them. The idea is: go where people are, go where they're already paying and interacting, and work with those restaurants to bring this information directly to the consumer. Gen Z and Gen X want to customize everything, and that's leaked into how everyone eats now. You have to know the nutrition, the allergens, what's in your food — I don't like onions, I don't like cilantro, okay, say so. It's your bowl, it's your food. We're trying to be the driving force that lets restaurants better serve their customers.