Photo by Elimende Inagella on Unsplash
Both the impressive and the stupid are true at the same time.
That’s what I keep coming back to after years building AI agents in a startup, and now as a product leader at Autodesk. Here’s how I think about what’s actually happening.
Follow the money
You can think the AI hype is overblown. I think so myself. But most of the capital in the world right now is betting that AI will fundamentally reshape most industries. The bubble will probably burst, and there will be winners and losers, just like every major technology wave. But that doesn’t change what capital does along the way: it creates innovation, draws talent, multiplies research, builds infrastructure, and creates irreversible shifts. We saw it with smartphones, cloud, EVs, SaaS. Each time, the people who dismissed the early, messy versions got caught off guard by how fast things compounded. AI is following that exact pattern, except faster, with even more money to fund the change.
What I learned building AI agents
When I was running my startup, we built AI agent systems. Every month brought a step change in capability. Something that worked horribly one week could be 10x better the next. That meant I couldn’t trust my own conclusions from the month before. I had to re-test constantly, and accept that what was true yesterday might not be true today.
I spent weeks designing around a model limitation I’d carefully mapped. By the time we shipped, the limitation was gone. The model had improved, and everything I’d built around the constraint was wasted architecture. That was the moment I realized my learnings had an expiration date ratio I’d never had to deal with before.
That experience rewired how I think about this. The people who struggle most with AI aren’t the ones who try it and find it lacking. They’re the ones who tried it once, even deeply, concluded it wasn’t good enough, and stopped checking every week. The rate of improvement is the thing to pay attention to, not any single snapshot.
Where it’s real and where it’s not (yet)
Tools like Lovable and Bolt aren’t for experienced engineers. They’re for the people who couldn’t build before. That’s a massive market: internal apps, CRUD apps, WordPress-level products. A lot of consultancies are going to struggle because of this. People are building real businesses on these tools alone.
For those who already understand what they need, tools like ChatGPT, Claude, Cursor, and Claude Code are removing the friction between intent and execution. The deeper your understanding, the more leverage you get.
Does AI solve all complex problems? No. Not yet. But something has fundamentally changed in the last few months. As a concrete example: I built a tool called Loom in about 60 minutes. It orchestrates a full AI product organization, every persona needed to build, market, and sell a product. It plans work from a vision doc, spins up parallel teams in isolated git worktrees, creates PRs, and has an AI reviewer that auto-merges what passes and escalates what doesn’t to me. It builds features for me while I do other work.
Is it sometimes dumb as hell? Yes. Does it make baffling product, design, and engineering mistakes? Yes. But has it permanently changed how I think about what’s possible? Absolutely. That tool would have been a multi-month side project a year ago, and impossible a few years before that. Now it’s built and I have a decent working product after half a week of vacation mostly spent with my kids and wife.
What I fundamentally believe
How we work as knowledge workers is changing at an extreme pace. You can choose to be at the frontier or follow from a distance. For me, being at the front is natural. But what you shouldn’t choose is to stop being curious. Dismissing it as “just hype” can be a defense mechanism. There’s also real transformation happening underneath.
We’ve always been abstractions away from natural language. From assembly to C, to Java, to JavaScript, to TypeScript, to “build me a tool that lets teams track projects.” That distance gets shorter every month now, and the complexity of what you can describe in natural language and actually get done grows every month too. All that knowledge of product taste, design patterns, architecture, and system thinking? It’s becoming your primary interface.
The people getting the most out of AI tools aren’t the ones writing the best prompts. They’re the ones who can look at what it produced and say “this won’t work, and here’s why.” The ones who understand the deeper abstractions and can direct it where they want to go. The tool does the building. The human designs the system it works within.
What you build matters too
This isn’t just about how you work. It’s about what you build. If you’re shipping products today without thinking about AI-native capabilities, agent-ready integrations, and machine-readable interfaces, your competitors are. Products that can’t be orchestrated by AI, that don’t expose their value through APIs and protocols like MCP, will feel like websites without mobile support felt in 2015. Not broken, just increasingly irrelevant. The window to build this in from the start is now. If your customer’s AI can’t use your product, they’ll find one it can.
What worries me
I’m not blindly optimistic here. I’ve been building and using AI systems for years. I get production-grade code out of AI regularly, but once you move beyond CRUD apps into complex, domain-specific territory, it’s still not completely there. It sometimes feels like wasted time, but I still think it’s worth it. The less training material that exists for your problem space, the worse it gets. The gap between what the frontier demos show and where most of us actually are day-to-day is real. If the industry oversells the promise without honestly bridging that gap, we’ll create cynicism instead of adoption.
I’m also genuinely concerned that we’ll become less personal, less human. Abstraction isn’t always good, and building on things you don’t fully understand carries real risks. The challenge is using AI to make us more human, not less.
Do I like the direction things are heading? I don’t even know. But things are moving there, and I’d rather shape the direction than watch it happen.
What I think matters now
The most important skill right now is the willingness to be wrong about what you concluded last week. That’s hard, especially if you’ve built a career on deep, stable expertise. If you’ve spent hours wrestling with AI-generated code that looked right but was subtly wrong, or watched it confidently produce nonsense, your frustration is completely valid. But the tool being bad today and the shift being real are two different things. The instinct is to form a view and hold it. With AI, that instinct will cost you.
If you’re skeptical, be honestly skeptical and engaged. Say where the tools fall short. That signal is more useful than either blind enthusiasm or silent dismissal. And if you’re not skeptical or excited but just unsure what this means for you: the work is changing shape, and some of what currently fills our days will be automated. But the parts that involve judgment, context, domain expertise, and navigating ambiguity are becoming more valuable, not less. AI makes the thinking more important, not less important.
I don’t have a tidy conclusion. This is messy, fast, and genuinely uncertain. Stay in the conversation, keep experimenting, and share what you find. What doesn’t work today might work next week.
Re-test. Every week. Stay uncomfortable.