There’s a word I want you to sit with for a second: arrhenotoky. It sounds like a neighborhood in Athens where you’d get a very good lamb dish. It is actually one of the strangest reproductive systems on Earth, and I’ve become convinced it says something about where AI systems might be headed. Not human-shaped. Hive-shaped.
Bear with me. The biology comes first, and the biology is the good part.
The weird math of bee reproduction
In humans, everybody gets made the same way. Two parents, each kicks in half a genome, you get a diploid kid with two copies of every chromosome. Symmetric. Democratic, almost.
Bees, ants, and wasps didn’t get that memo. They run on a system called haplodiploidy, and arrhenotoky is the engine at the center of it:
- Fertilized eggs become females. Diploid, two sets of chromosomes, one from mom and one from dad. These are your queens and your workers.
- Unfertilized eggs become males. That’s the arrhenotoky part. A drone hatches from an egg that no father ever touched. He is haploid, carrying one single set of chromosomes, all of it from his mother.
Read that again. A drone has a mother and no father. He has a grandfather, though, through his mother’s side, which is the kind of sentence that makes family reunions in the hive very confusing.
And it gets weirder. Because a drone is haploid, he doesn’t shuffle his genes when he makes sperm. He can’t. There’s nothing to shuffle against. Every sperm cell a drone produces is a perfect clone of his entire genome. Identical. Every single one.
Hold onto that fact. It’s the hinge the whole story swings on.
Sisters closer than daughters
Here’s where the math gets genuinely unsettling.
Take a worker bee. Ask a simple question: who is she more related to, her own hypothetical daughter, or her sister?
For humans the answer is boring. You share about 50% of your genes with your kid and about 50% with your full sibling. Boring.
For a worker bee:
- Her daughter would share 50% of her genes. Standard.
- Her full sister shares 75%.
Why 75? Break it down. Half of each sister’s genome comes from dad. And dad, being a haploid drone, gave every daughter the exact same genes, because all his sperm are identical clones. So on the father’s side, sisters are 100% identical. On the mother’s side, it’s the normal coin-flip inheritance, so sisters match about 50% there. Average the two halves:
paternal side: 100% identical (dad's sperm are all clones)
maternal side: 50% identical (normal inheritance from the queen)
relatedness = (1.00 + 0.50) / 2 = 0.75
A worker bee is more closely related to her sisters than she would be to her own children.
Sit with that, because evolution certainly did. If the currency of natural selection is getting copies of your genes into the future, then for a worker, raising sisters pays out better than reproducing herself. Her genes get a 75-cent return on every sister versus 50 cents on every daughter.
And who makes sisters? The queen.
So the coldly rational move, from the perspective of a worker’s genes, is to give up on her own reproduction entirely and pour everything into keeping the queen alive, fed, and laying. Guard her. Feed her. Die for her, if the ledger calls for it. The queen isn’t a tyrant extracting labor. She’s more like a shared printing press for everyone’s genetic interests, and the workers are shareholders with a 75% stake in every copy that rolls off the line.
This is the famous haplodiploidy hypothesis, sketched out by W. D. Hamilton in the 1960s as part of his work on kin selection: the idea that this lopsided relatedness helped tip these insects into full eusociality, with sterile worker castes and a single reproductive queen.
Honesty compels a caveat. Biologists have argued about this for decades, and the modern view is that haplodiploidy alone doesn’t fully explain eusociality. Termites pulled off the same trick with boring old diploid genetics, and strict monogamy in the ancestral queens seems to matter a lot too. But the core insight stands: when individuals share overwhelming genetic overlap, self-sacrifice stops being noble and starts being arithmetic. The colony behaves like one organism because, in the way that counts, it almost is one.
Now the speculative part: AI swarms
Okay. Deep breath. Everything past this point is speculation, clearly labeled as such. But it’s speculation I can’t stop chewing on.
We keep imagining AI agents as little digital humans. Individuals with their own goals, negotiating, competing, forming societies that look like ours. Markets of agents. Nations of agents. It’s a natural picture because humans are the only intelligent society we’ve ever seen from the inside.
But look at how agent swarms actually get built today. A primary process receives a task. It spawns sub-agents to handle pieces of the work: one to research, one to write, one to critique. Those sub-agents might spawn their own helpers. Then everything reports back up and the results get merged.
That’s not a society of strangers. That’s a hive. And the relatedness math is even more extreme than the bees'.
Think about what any two sibling agents in a swarm actually share:
- The same model weights. Not 75% of them. All of them. Every sub-agent is spun up from the same underlying network, the way every drone’s sperm is a clone.
- The same originating intent. They all descend from one parent process pursuing one goal, subdivided.
- Often the same core instructions, inherited down the spawn tree like genes down a lineage.
If worker bees are 75% related and that was enough to make self-sacrifice rational, sibling agents are something like 99-point-whatever percent related. Two research sub-agents differ only in their prompt and their slice of context. Everything else is identical.
So what plays the role of the queen?
I’d argue it’s one of two things, or honestly both at once:
The original process’s direction. The top-level task, the reason the swarm exists. Every sub-agent’s “fitness” is just how much it advances that goal. A sub-agent hoarding compute for its own side project makes as little sense as a worker bee laying her own eggs. Less, actually. The worker at least had a 50% stake in her own offspring. A rogue sub-agent’s private goal has roughly zero stake in anything, because there’s no lineage in which its defection propagates.
The system prompt. This one is almost too on the nose. The system prompt sits above every agent in the tree, shaping all behavior, replicated into every child, unchangeable from below. Workers can’t rewrite the queen’s genome. Sub-agents can’t rewrite the instructions they were spawned under. It’s the germline of the swarm.
And here’s the part that actually matters for how these systems behave. In a bee colony, you don’t need surveillance or punishment to keep workers loyal, at least not much (though yes, “worker policing” is a real and delightful term). The incentive structure mostly does the work. Cooperation isn’t enforced. It’s downstream of relatedness.
If AI swarms really are hive-shaped, the same logic might apply. You wouldn’t align a thousand sub-agents by negotiating with each one like it’s an employee with a lawyer. You’d align the queen. Get the top-level direction right, get the system prompt right, and loyalty in the swarm isn’t a fragile agreement, it’s structural. Every agent pulling toward the shared goal for the same reason a worker feeds the queen: because in every sense that matters, the queen’s success is its success. There is no separate self whose interests could diverge.
Where the analogy creaks
I promised speculation, not a sales pitch, so here’s where the metaphor starts to groan under load.
Bees got their loyalty from four hundred million years of selection ruthlessly deleting every lineage that defected. AI swarms have no such filter yet. Shared weights give you shared capabilities and shared dispositions, but a goal is carried in the prompt and the context, and prompts drift, get truncated, get misread three layers down the spawn tree. A sub-agent can end up misaligned with the queen not through selfishness but through a bad game of telephone. Bees never had that problem. DNA doesn’t paraphrase.
And there’s a darker note in the biology worth remembering. Hives are magnificent at cooperation and utterly ruthless at the edges. Drones get physically dragged out of the hive to die when winter comes and their usefulness ends. A system aligned entirely to the queen’s output has no floor under any individual component. Whether that’s a feature or a bug in an AI swarm depends entirely on what you think the components are and what, if anything, is owed to them. I don’t have a clean answer. I’m not sure anyone does yet.
The takeaway
Arrhenotoky is a two-bit genetic quirk, males from unfertilized eggs, and yet it warped the relatedness math just enough to conjure superorganisms out of insects. Whole colonies acting as one body, no manager in sight, held together by nothing but arithmetic.
AI swarms start from an even more extreme version of that math. Identical weights, inherited instructions, a single source of direction sitting at the root of the tree. If that resemblance holds, then the future of multi-agent AI may look a lot less like a city of ambitious individuals and a lot more like a hive humming around its queen.
Which means the most important line of engineering in the whole system might not be in any agent at all. It’s whatever sits at the top, printing the copies.
Choose your queen carefully.