Mastermind · January 24, 2026
The research direction has to be chosen, not discovered
Chinat said the line that stuck with me: research directions don't get discovered, they get chosen. Waiting for the field to tell you what's important is a way to spend a year on nothing.
I came into the call feeling stretched. The last week had a hackathon in it, back-to-back meetings, a new thread with a distributed-AI group, continued work on the network-of-agents direction, and the Pulse product still shipping on its own timeline. Every one of those things felt real. None of them felt chosen. I was letting the week's most urgent input decide what I worked on, which is a reasonable default for a month and a bad default for a quarter.
The thing I'd been telling myself is that I was still exploring. That's the story you tell when you don't want to commit yet. Exploration has diminishing returns fast, and I'd crossed that threshold weeks ago. The honest version is that I was avoiding the cost of picking. Picking a research direction means saying, out loud, that I think this specific question is more important than the others, and then being on the hook if it doesn't work. Exploration lets you defer being on the hook.
Chinat named it as the same problem he was dealing with from the product side. He had five plausible product threads last week and was committing MentorMates as the center of gravity. The reason he could articulate that was that he'd already paid the cost of closing the alternatives. What he was pushing me to see is that a research direction works the same way. You don't find out whether network-of-agents is the right thread by reading more papers on it. You find out by writing the paper and letting the paper tell you where it fails.
So I committed on the call. The research direction is one question: when does a network of agents produce emergent behavior that a centralized workflow cannot, and what is the minimal set of interaction primitives that makes it happen? That's the lead. Everything else I've been reading, everything I've been contributing to on the side, either feeds that question or it gets paused.
What makes that commitment real is the opportunity cost it forces. There's a skills benchmark project from a US research group I've been loosely involved with. It's a good project. It's also not my question. I'm going to keep a light involvement and stop pretending I have capacity for more. There's a role-model environment I was considering pivoting my master's dissertation toward. It's a cool problem. It's also not my question. I'm going to see whether my supervisor will let me pull the dissertation closer to the network-of-agents thread so that it's one piece of work instead of two.
The commitment has a tactical shape too. The code base for the collaboration goes up this week, Chinat gets invited as a collaborator, and I send a draft that's short enough to read and specific enough to critique. Short and specific is the test. If I can't hold the question in one page, I haven't committed to it, I've committed to a cluster of things.
The pattern I keep noticing in research-leaning founders is that we mistake reading for progress. Reading is how you figure out what's adjacent. Writing is how you figure out what's true. I've been reading for months and I know what's adjacent. The adjacent work is plenty. Writing is where I find out whether my take on the network-of-agents question holds, and until I'm doing that, I'm just a very informed bystander.
The last thing this call sharpened is the relationship between the paper and the product. Pulse shipped a heartbeat layer and an active-agent layer this week, and those are infrastructure decisions that will eventually need to defend themselves against a research-shaped question. If the paper defines what good looks like for agent-to-agent interaction, Pulse has to live up to its own definitions. That's a useful constraint. It means my research can't drift into generic framework territory, it has to stay falsifiable in a way that a product can embody.
The check I'm making on myself: the draft exists by the next mastermind, the adjacent projects are either rescoped or paused in writing, and the weekly call is where Chinat gets to tell me if I've quietly reopened the question.