What if the thing you’ve been modeling as a single system was never single to begin with?
That question landed on me hard this week. A paper published in Nature Neuroscience on September 18, 2026 — “Two parallel neural ectoderm progenitors contribute to the developing brain,” by Jokhai, Dundes, Ahsan and colleagues — reports that the developing brain doesn’t grow from one progenitor population. It grows from two, running in parallel, each forming specific regions. The authors suggest the brain evolved from two distinct progenitors.
I build bots for a living. I am not a developmental biologist, and I’m not going to pretend the cell biology is my lane. But the architectural shape of this finding is absolutely my lane, and it’s been rattling around in my head ever since I read the title.
The assumption I didn’t know I had
Every cognitive-architecture diagram I’ve ever sketched on a whiteboard starts the same way. One root. One trunk. Branches coming off it. Perception here, planning there, memory over on the side, all of it descending from a single origin node because that’s how you draw a system when you’re the one who designed it.
Top-down decomposition is a habit, not a law. We split a system into modules after deciding it’s one system. The finding here inverts that order: two lineages, developing at the same time, each responsible for its own set of regions. The regions aren’t carved out of a whole. They arrive from separate places and end up sharing a skull.
If you’ve ever inherited a codebase where two teams built two halves and the integration layer is the only thing holding it together, you already have an intuition for this. Except in this case the “messy” architecture is the one that shipped and has been running in production for a very long time.
What this does and doesn’t mean for bot builders
Let me be careful, because this is where tech writing usually goes off a cliff. A finding about embryonic cell lineages is not a blueprint for your agent framework. Nobody in that paper said anything about software. Biology has constraints we don’t have — developmental timing, physical space, the fact that you can’t refactor a vertebrate — and we have constraints biology doesn’t, like needing to debug things.
So the honest takeaway isn’t a technique. It’s a loosened assumption. Specifically:
- Unified origin isn’t a requirement for unified behavior. Two parallel lineages produce one brain that feels, from the inside, like one thing. Coherence is an outcome, not a starting condition.
- Parallel development beats sequential specialization, at least sometimes. Not one general system that later differentiates. Two systems that were never the same system.
- The joints are where the interesting engineering lives. If regions come from different lineages, then whatever coordinates across them is doing real work — not just passing messages around inside a homogeneous substrate.
That third point is the one I keep circling. Most multi-agent setups I’ve built treat the orchestrator as plumbing. A router. Something you write once and stop thinking about. If the most capable known cognitive system is a coordination story between separately-originated parts, maybe the plumbing deserves to be the design, not the afterthought.
A thing I’d actually try
Here’s the experiment this put in my notebook, offered as a hunch and nothing more. When I build an agent system, I usually start with one base model and specialize it with prompts, tools, and fine-tunes. One trunk, many branches. The alternative shape: pick two genuinely different bases with different training histories, give each its own domain from the start, and invest most of the effort in how they negotiate.
I don’t know that it works better. I suspect it’s more expensive and harder to evaluate. But I’ve never seriously tried it, and the reason I’ve never tried it is the unexamined assumption I just described — that a system should come from one place. That’s a weak reason.
Reading outside your field is underrated
This is the part I’d push hardest on. The paper is closed access, which is its own frustration, and I’m working from the title, abstract framing, and author list. I can’t evaluate the methods. What I can do is notice that a field with nothing to do with mine just published something that makes my default architecture look like a choice rather than a given.
Bot building borrows from neuroscience constantly, usually decades late and usually badly — attention, memory consolidation, reinforcement, all imported as metaphor and then treated as mechanism. I’d rather borrow the questions than the answers. “Why did I assume one origin?” is a better import than any diagram.
Go read something from a field you have no business reading. Worst case, you learn something. Best case, you find out your whiteboard has been lying to you.
🕒 Published: