Hard or Not Working?

I used to be really embarrassed about pivoting. Actually, I still am a little bit. Every time I change what I’m working on, there is a part of me that is ashamed. Everyone is going to think I failed again. It is especially hard when you have been talking publicly about what you are building. You tell people about it. You post about it. You get excited. You convince yourself you have finally found the thing. Then, a few weeks or months later, you are explaining why you are doing something different.

There is no way to make that feel completely normal. but I think I’ve finally stopped thinking that pivoting itself is the failure. The failure would be knowing something isn’t working and continuing to work on it because I’m too embarrassed to change my mind. I get super excited about ideas... probably too excited. I can convince myself of a very large future very quickly. But weirdly, I’m also not that attached to the actual idea. I’m much more attached to figuring out what is true.

If I learn something that makes my original thesis wrong, I want to know. If someone tells me something that completely changes the way I’m thinking about the problem, I want to pull on that thread. If we build something and people use it but there is no coherent business underneath it, I can’t really unsee that.

I think that has made the past few months emotionally chaotic, but also incredibly educational. There is a difference between something being hard and something not working. Startups are obviously supposed to be hard. That phrase gets repeated so often that it can almost become dangerous. Because you can use “startups are hard” to justify basically anything.

No customers? Startups are hard.

Nobody will pay? Startups are hard.

You don’t know who the customer is? Startups are hard.

You’ve spent a month inventing increasingly complicated explanations for how this eventually becomes a business? Well, startups are hard.

At some point, though, you have to distinguish perseverance from refusing to update your beliefs. For me, one of the biggest signals was monetization. With PlateTale, we moved incredibly fast. That was actually one of the best things we did we launched things constantly.

Launch. Learn. Change something. Launch again.

We started with food. At one point we were operating a marketplace for home-cooked meals. We got real transactions, repeat customers, chefs, logistics, payments, pickup, delivery—the whole thing. Then we realized that the more interesting problem wasn’t necessarily access to food. It was deciding what you actually wanted to eat. So we moved toward personalization. That became this much bigger technical obsession for me could you build an AI that actually learns a person well enough to make decisions for them? Not “you liked sushi once, here are twelve sushi restaurants.” Actually learning you.

I still think that problem is fascinating.

But fascination is not a business model.

We kept trying different ways of making the business work. Consumer subscriptions. Commerce. Commissions. Brand partnerships. B2B. Different ways of sitting somewhere between recommendation and transaction. and while there were technically ways to make money. that wasn’t really the problem.

The problem was that the ways that seemed most likely to make meaningful money increasingly pulled us away from the thing we actually cared about building. At some point, if the only economically attractive version of your company destroys the reason you wanted to build the company, that feels like important information. And when you have spent weeks trying to construct a revenue strategy and every answer requires five more assumptions before it starts making sense, that is also information. It doesn’t mean the idea is objectively bad. Somebody else might build an enormous company in that exact market. It means I couldn’t see the path clearly enough to justify continuing to spend years of my life on it.

That is a very different thing from quitting because something got difficult.

The useful part is what happens while you’re wrong. The weird thing about pivoting is that almost none of the work disappears. You think you are building one company, but the entire time you are accidentally collecting information about the next one.

Trying to personalize food recommendations forced me to think seriously about agents. If you actually know what someone wants, the obvious next question is: why are you still just recommending things to them? Why doesn’t the system just do it?

If an agent knows the restaurant I like, knows what I usually order, knows my budget, knows that today I want something healthy, why am I opening five apps and clicking around? ok, build an agent. Then you immediately discover that building the “intelligence” is only part of the problem. Because now your very smart agent has to use the internet, which was not built for agents. This sounds stupidly obvious once you see it.

Humans are unbelievably good at navigating bad websites. We can look at a giant hero image, ignore three pop-ups, understand that “experience our collection” probably means “shop,” find a tiny dropdown, infer what a badly labeled button does, notice that a hotel’s cancellation policy is hidden three pages away, and eventually accomplish what we wanted.

An agent has to somehow reconstruct all of that. And when it gets it wrong, sometimes it doesn’t even know it got it wrong.

That was one of the threads that pulled me away from thinking only about personalization and toward thinking much more broadly about the infrastructure around agents.

The models are becoming insanely capable but the environment they have to operate in is not.

We really do not know how to use agents yet.

I keep coming back to this. People do not know how to use agents. I don’t mean that people don’t know how to type a prompt into ChatGPT. I mean we have these increasingly autonomous systems and almost none of the infrastructure surrounding them feels settled. How should they remember things?
How much context should they have? How much money can it spend? What information is it accidentally revealing to another model or service? ...


And then there is the other side of this: the businesses the agents are interacting with. That part has become especially interesting to me. For the last twenty years, companies have spent enormous amounts of time controlling how they appear on the internet.

Their website, SEO, product descriptions, brand, checkout flow, recommendations, ads.

All of that assumes a human is the one arriving.

But increasingly, the first thing interacting with your business might not be a person. It might be their agent. and suddenly your beautiful website might be completely useless.

The product information the human understands immediately might be ambiguous to the agent. The thing that makes your hotel special might be communicated almost entirely through photography. Your policies might be scattered across six pages. The difference between two products might be obvious visually but almost impossible to extract structurally. The agent may misunderstand something, summarize you incorrectly, recommend a competitor, or fail halfway through a transaction.

## Every conversation has made the problem bigger

One thing I have learned about finding startup ideas is that I don’t think I need to pick a “problem space” and force myself to stay inside it. For a while I thought I did.

Like: choose food. Choose personalization. Choose cybersecurity. Choose developer tools. Whatever.

But almost every interesting thing I’ve learned recently has come from crossing those boundaries. I’ll talk to someone about agent harnesses and suddenly I’m thinking about reliability. I’ll build an agent myself and watch it do something completely idiotic. I used to find this lack of a neat category stressful, now I think it might actually be how I find things.

Every conversation gives me another piece. You do not necessarily need to wake up knowing, “I will spend the next ten years working on cybersecurity for autonomous agents.” You can just find a problem that keeps bothering you enough that you cannot stop pulling on it.

And right now, the thing that bothers me is the gap between how capable agents are becoming and how unprepared everything around them is.

The intelligence is getting better much faster than everything around it. The web, businesses, privacy, interfaces, even users themselves, are still built for a world where humans click the buttons.

Something about that has to change.

So yes, we pivoted. Again.

I wish I could make this sound more cinematic. Like we discovered one brilliant insight on a Tuesday afternoon, deleted the old codebase, and suddenly saw the future. That is not what happened. It was much messier.
A lot of reading, building, realizing I was wrong, tiny observations that eventually started pointing in roughly the same direction.

I still don’t know exactly what the final company looks like. I’m much more okay saying that now.

What I do know is that I want to keep working on this problem.

I want to understand where agents fail in the real world. I want to understand what websites look like when they are designed not only for humans, but for humans and the agents acting for them.

Maybe this version will change too.

That sentence used to terrify me.

Now I think: good.

If we learn something important enough that we have to change again, we should.

The point is not to prove that the first thing you thought of was right. The point is to get closer to the truth every time you build something. I think being a founder is partially becoming comfortable with publicly changing your mind. In a way isa science experiment. You make a hypothesis.You test it. Observe results. Try again.

I would rather be embarrassed about pivoting than spend three years defending an idea I already know I don’t believe in.

So this is where I am right now.

Still building.

Still probably overly excited.

Still changing my mind constantly.

But much less embarrassed about it. byeeeee:))))