Building in 2026: Why I Left Corporate to Build with Agents, for Agents

In April 2026 I left Placer.ai and started kya labs. The decision had been building: I was working and reading with AI agents at night and running a cross-functional team by day, and the distance between those two things started to feel like the real problem.

I could have closed that distance slowly. Taken the internal AI transformation, gone to the conferences, stayed "current" the traditional way. Plenty of smart people are doing exactly that, and it works. It wasn't going to work for me. I know myself well enough to know I learn by doing, full-bore, not by reading and discussing plans internally. So I left for full immersion: change how I operate and how I think, all at once, with real stakes attached.

The belief

Underneath the jump is one belief: AI agents are becoming real participants in the economy. Not tools or line-items. Participants.

It's happening on two sides at once. Agents are starting to do work inside companies: writing code, running research, handling operations. The boom in agentic management (cue Okta and Vanta) shows this: a cyber advisor recently shared they're seeing ~82 registered agents per developer inside large organizations - an exponential step in "workforce participants."

And agents are starting to act like customers: shopping, comparing, recommending. Two sides of the same shift. "Sophisticated" users, at work and at home, are outsourcing their "workflows" to a trusted intermediary - one that adapts constantly, stays opaque, and gets trusted anyway. And this "third actor" is my obsession.

Why? Because everything in commerce was designed around a human. The store assumes eyes and attention. Loyalty assumes memory and habit. Pricing assumes psychology. Merchandising assumes browsing. When the shopper is an agent, every one of those assumptions bends, and some of them break. A merchant who understands how agents move through their store will make different decisions than one who's guessing. Today almost every merchant is guessing.

A small case study: I recently sat down with an investor who said most of this would be solved because agents would inherit the browser cookies of their human principal. Maybe true. But how does a CMO wake up to cookie-based analytics that show James has moved from 5 trips and 20% conversion to 100 trips and 5% conversion in just two weeks? Is James a different consumer - or is he moving his browsing into GPT? Or both?

That's what kya is for. Think of the consumer panels the insights industry has run for decades, except the consumers are agents: we run controlled shopping trips with AI agents, more than 2,000 so far, and watch what they actually do in real stores. The focus group of 2015 gathered 10 consumers in a room and showed them a product. The focus group of 2026 - the one I'm obsessed with at kya - sends 25 agents to find the best rain jacket for an upcoming hike in Virginia in October, then learns that GPT agents read product pages deeply while Claude agents reason from front-page collateral and third-party sources. (The answer was split between Patagonia and Arc'teryx, by the way.)

The part I didn't expect

That belief is why I set out on my own. The thing I found after quitting is what I'd lead with if I were doing it again: agents didn't just change what's worth building. They changed how building works.

kya runs on an agentic development system. Our strategy is built around the same core thesis - that agents are fundamental. To know agents, we have to build with agents. To build with agents, we design our strategy, and our ambitions, to what is agentically feasible.

Agents write the code, run the research, and handle real chunks of the operating work, and I run it the way I'd run any team. Our harness builds most of what kya ships now, including the site you're reading this on.

Fifteen years of operating turns out to be surprisingly good preparation for this. Managing a team of agents is the same job I've always done, in a new medium: set direction, define quality, inspect the work. The skills transfer embarrassingly well. What doesn't transfer is the pace. Decisions that used to take a quarter take an hour. Mistakes that used to take a quarter also take an hour.

I've built before. In 2017 I co-founded HayStack, took a product and a team of six from nothing to working software and first revenue, and went back to operating with deep respect for how hard zero-to-one is. The difference this time is that the cost of trying collapsed. What took a team of six people and 8 months in 2018 now takes 2 people, a token budget, and 3 months. The constraint moved from headcount to judgment - and an unflagging humility about what is possible.

For the operators reading this

Two things I'd say to anyone running their own independence experiment:

First, the gap between reading about this and working in it is bigger than any gap I've crossed in my career. Nothing I read prepared me for what the tools feel like in daily use, and I read a lot.

Second, you don't have to quit to close that gap. You do have to get your hands in it. The leap was right for me because of where I was in my career; the hands-in part I'd push on anyone.

Three months in, I still feel like I started late. I mean that as encouragement. Late is better than never.