Xisen · August 25, 2026

We Changed Our North Star

Four weeks ago, on this same mastermind call, my north star was three paying teams. It was a good north star by every standard playbook: revenue, real users, willingness to pay. This week Chinat read it back to me from his notes and asked how it went, and I told him the truth. We changed it.

The new north star is this: a realistic chain of agent delegation that completes one genuinely hard task. Concretely: hand a group of agents a one-line prompt and have them rebuild a working program, thousands of lines, work flowing from one agent to the next, on a task where a single agent stalls out. Not three customers. Not a launch. One demonstration that interdependent agents can accomplish something a single agent cannot.

On paper this looks like a retreat from commerce into research. I want to argue it's the opposite: it's choosing the part of the problem where our advantage is real, the coordination, permissions, and shared state between agents.

Here is the reasoning. Over the past two months of working on this full time, I kept running into the same wall. In agent products, everyone is building the same surface. Assistants, chief-of-staff bots, delegation features. The large platforms ship them as line items. On the call, Chinat showed me a major AI company's new bot product that already does agent delegation inside group chats, priced and shipped. If we compete on product surface, we are competing with teams that have a hundred times our distribution, and our differentiation evaporates the moment any of them wakes up.

So I'm choosing to compete on research instead. That sounds like avoiding competition, and in a sense it is. But here's the thing I said to Chinat that I keep coming back to: you can't actually cheat this way. If you compete on product, you need a good product. If you compete on research, you need good research. At the end of the day, you still need good work. The only choice you get is which kind of good work plays to your strengths.

Our strengths are the unglamorous infrastructure layers. We have spent months on SharedOS, which handles the problem nobody finds sexy until it bites them: when one person's agent asks to read another person's files, who authorizes that, and how? Authorization, multi-owner management, agent-to-agent permissions. There are now ten independent researchers running their own experiments on this infrastructure, studying things like agent negotiation and social simulation, and we run a weekly research meeting around it. That pull, researchers choosing our infra because it's the only way to run their experiments, is a different kind of signal than a paying customer, but it is not a weaker one.

The north star has a measuring stick, which matters, because "we do research now" is how startups dissolve into vibes. There's a benchmark built by researchers at top labs where a model gets a one-line prompt and has to rebuild an entire program from scratch. Thousands of lines. Real software, not toy exercises. Single agents do poorly on it. Chinat immediately pushed on contamination, since the target programs are public repositories the models trained on. It's a fair push and it isn't fully settled. The benchmark team measures direct memorization, and rebuilding thousands of lines of coordinated structure is a much harder thing to regurgitate than a snippet, but the question stays open and anyone auditing our results should poke at it. If a chain of our agents climbs that leaderboard where single agents stall, that is the demonstration. It is legible to researchers, to engineers, and yes, to investors, without us having to say a word about vision.

What surprised me is how much calmer the new north star makes me. "Three paying teams" pointed at a dozen activities at once: sales calls, onboarding, pricing experiments, support. Every day offered ten reasonable things to do, which in practice meant every day was a negotiation with myself. "One chain of agents completes one hard task" points at exactly two priorities this week: finish the first draft of our technical report and get it in front of my advisors, and stand up a proper cloud server so experiments stop running on my laptop. That's the whole week. I can hold it in my head in the shower.

There's a personal parallel I'll admit to. My life right now has more uncertainty in it than any year since I was eighteen, most of it outside my control. The thing I've reached for, and I noticed this only after rereading Paul Graham's Hackers and Painters this month, is a small set of invariants. I read one excellent codebase every two days. I read one good paper every two days. Whatever else moves, those don't. The north star change is the company-scale version of the same move: when the environment is uncertain, anchor on the thing you can verify with your own hands.

The honest risk is that this becomes a hiding place. Research timelines are forgiving in a way markets are not, and "the demo isn't ready yet" can absorb infinite months. So the check is short: by next week's call, I owe Chinat a chain of agents completing a single real task, however modest. Not a plan for one. One that ran.

← Back to all posts