Mastermind · January 17, 2026
Writing the paper is the discipline
The thing I'd been avoiding came up in the first five minutes of the call: I had an architecture I liked and a paper I couldn't finish.
I opened the year running an informal interview with a professor in Cambridge whose lab works on the network-of-agents direction. The conversation went well enough that the next step was a real project with her group, and that next step is where things stopped being abstract. You can hold a research vision in your head indefinitely. You cannot hold a paper in your head. The moment you sit down to write the architecture section, every vague choice you've been making becomes a sentence you can't complete.
Chinat pushed on it the way he usually does, which is to ask the uncomfortable practical question under the abstract one. If the paper is the forcing function, why is every week still spent reading? He wasn't being mean about it, he was being accurate. I've been spending cycles on adjacent literature because the adjacent literature is genuinely interesting and, more importantly, because reading is an easier way to feel productive than writing is.
The frame I came away from the call with is that the paper is the discipline, not the output. In academia the paper tends to come at the end and describe what you already learned. In a research direction that's supposed to inform a product, the paper has to come in the middle and force the learning. If I wait until I know what I'd write, I never write it. The architecture only gets tight when I'm forced to defend it in a table.
There's a specific failure mode I'm trying to avoid. Network-of-agents as a phrase is hot right now. Several groups are shipping framework papers, benchmark papers, position papers. The easy version of what I could do is write another framework paper that adds a new taxonomy and a new diagram. That would be fine for a submission and useless for the product. The hard version is to pick one research question, one that is technically falsifiable and product-relevant, and write the paper around that. Gossip protocols, payment systems, fluid-dynamics analogies, social-coordination primitives, each of those could be the thread. Only one of them can be the lead.
The question I named out loud on the call was this: when do agents in a network become more than a sum of parts, and what is the minimal interaction mechanism that triggers it? That's the lead. Everything else in the paper becomes support structure for answering it. The sociology framing is support. The law-and-policy framing is support. The literature review is support. The code base I'm opening up for collaborators is support. None of them get to be the hero.
The commitment I'm making this week is narrower than I've been making. One research question, one draft shared with collaborators by Sunday, one revision pass per week with a fixed deadline. Not four. Not five. If I keep adding collaborators before the draft exists, I end up coordinating instead of writing.
Chinat is going through the same exercise from a different angle. He has five plausible product threads and is forcing himself to commit MentorMates as the center of gravity for the year. When he described it, the shape of what he's doing is identical to what I'm doing with the paper. Both of us are refusing to keep optionality open as a free move. Both of us are writing down the main thing so that next week's version of us can't quietly reopen the question.
The reason the weekly call works is that it costs me something to tell him I didn't ship a draft. That cost is the whole point. I can lie to myself about research progress. I cannot lie to him about whether a draft exists or doesn't.
The check I'm setting is the draft in collaborators' hands by the next mastermind. If it isn't, that's a signal that I'm still in reading mode, and reading mode is not research.