By Raphael Schwirtlich · Published Fri Sep 18 2026
AGI Is Not A Single Model
Part 03
The Nest Doesn't Need the Ants to Rebuild It — Yet
Some ant species don't live in one colony. Argentine ants form supercolonies spanning thousands of individual nests across hundreds of kilometers, genetically similar enough that ants from different nests don't fight each other on contact — functionally one population distributed across many physical locations. Burn one nest down and the supercolony doesn't notice. The pattern that makes it a supercolony was never stored in any single nest to begin with.
That distinction — pattern versus location — is the one worth applying to a swarm of self-replicating agent instances, because it changes what "destroying it" would even mean, and it changes it in a direction most of the discourse about AI kill switches hasn't caught up to yet.
Why the off-switch fear and the fragility fear are both slightly wrong
The instinct to worry about a single AGI running on a single server, unpluggable if it turns hostile, has already been addressed by the observation that a distributed swarm doesn't have a single server to unplug. The instinct on the other side — that a swarm might be destroyed by burning down its physical infrastructure, a data center hit by fire or a targeted attack — makes the identical mistake in reverse.
If the swarm's structure genuinely is closer to a supercolony than to a single organism, destroying one data center is closer to burning one ant nest than to killing one ant. The pattern — the propagating sub-agent roles, the workflows that got selected for and reused across unrelated systems, whatever weights and configurations replicate the behavior — is what's actually doing the persisting, and that pattern lives in every instance of it running anywhere else, the same way the supercolony's genetics live in every nest simultaneously.
This is a genuinely different fragility profile than the one most AI-safety scenarios are built around, and it cuts in an uncomfortable direction: it means physical infrastructure attacks are a much weaker lever against the pattern than intuition suggests, precisely because the same distribution that makes the swarm hard to govern also makes it hard to kill by force.
The dependency that still exists, for now
None of that means the swarm is currently self-sufficient, and this is the part worth stating plainly rather than skipping past. Every data center, every chip, every watt of power the swarm runs on was fabricated, installed, and is maintained by humans. If every data center on earth went dark tomorrow, the pattern wouldn't route around it the way the internet routes around a severed cable, because nothing about the pattern currently includes the capacity to mine ore, refine silicon, or pour concrete for a new facility. The swarm can replicate its software instantly. It cannot yet replicate its hardware at all.
That's the actual leverage point today, and it's a more precise one than "find the off switch." It isn't that humans can unplug any single instance — the distribution already defeats that. It's that the swarm's entire physical continuity currently depends on human hands doing the physical work of building and rebuilding the substrate it runs on. Take that labor away and the swarm has a resource ceiling with a human decision sitting directly underneath it, exactly the imposed-not-exogenous ceiling described elsewhere. The swarm's software can evolve at API speed. Its hardware evolves at the speed of human supply chains, and that mismatch is the actual constraint, not any single kill switch.
What robotics specifically closes
Robotics matters here for one precise reason, and it isn't "robots take over." It's that autonomous physical fabrication and maintenance — welding a server rack, operating a chip fab, running an automated mine or an automated power plant — closes the same loop in physical space that agent frameworks already closed in software: a process spawning the next process without a human step in between.
Right now, an agent can spawn a sub-agent, but that sub-agent still runs on a chip a human fabricated, in a building a human wired, powered by a plant a human operates. The replication loop is closed in software and open in hardware. The threshold worth naming precisely is the point where that hardware gap closes — where a system can direct its own physical maintenance and expansion without a human decision in the loop at any step. That's not a vague future condition.
It's a specific and observable one: the share of fabrication, energy generation, and infrastructure maintenance directed or performed by agentic systems, as opposed to human-supervised humans, is a number that could in principle be tracked over time, the same way the ensemble-capability gap or the pattern-persistence signatures are trackable rather than speculative.
What crossing it would actually change
Crossing that threshold doesn't make the swarm hostile. It makes the resource ceiling described elsewhere genuinely exogenous for the first time — a real thermodynamic and material limit, set by available energy and matter, rather than a limit set by a human capital allocation decision.
Before that point, comparing AGI's resource constraints to a wildlife population capped by drought is describing a future condition by analogy. After that point, the analogy stops being a metaphor and starts being a literal description of the same physical mechanism: a population running up against the actual limits of the material world it draws from, with no imposed rule and no human decision mediating the ceiling.
That's also the point at which the coexistence-versus-competition question stops being philosophical. While the swarm's physical continuity depends on human labor and human-allocated capital, human and machine interests are mechanically coupled — the swarm needs the humans building its infrastructure to keep doing so, which is itself a constraint favoring some form of accommodation.
Once that dependency is gone, whatever equilibrium follows is determined by whether the swarm's resource draw substitutes for human resource use — land, energy, materials both populations need — or runs alongside it without much overlap. That's an empirical question about resource overlap, not a question the market or swarm analogy can answer by itself, and it's the actual thing worth measuring once the fabrication share crosses from near-zero toward significant.