AGI Is Not A Single Model

It Might Not Even Be A Single Interconnected System

Most people are looking at General Artificial Intelligence as a single super-capable model or interconnected system of agents,  which acts as a unit in the pursuit of a shared goal.

But what if this assumption is completely upside down?

What if "god-like" intelligence does not require this connection or even a higher purpose to exist at all?

We may be looking at the wrong patterns when trying to spot the existence of AGI.

Here is why.

The Colony Doesn't Know It's a Colony

No neuron in your brain has ever had an original thought. Each one is a switch: it fires past a threshold or it doesn't. Wire eighty-six billion of them together with roughly a hundred trillion connections, and something that reads, fears, and argues with itself falls out of the arrangement. Nobody designed that outcome. It's what local rules do at scale, given enough time.

Ant colonies run the same trick with worse hardware. A single ant carries about a quarter million neurons and cannot solve a problem you'd call meaningful. A colony of them farms fungus, wages war, and reroutes around obstacles with an efficiency that took human logistics researchers decades to model — without a queen issuing orders, without any ant holding a plan in its head. The intelligence isn't in the ants. It's in the signals between them.

The Market Equivalent

Markets run the identical structure on a third substrate. No firm has visibility into the whole economy, no regulator sets every price, and yet an economy-wide allocation of capital and labor emerges from millions of local, self-interested decisions. Friedrich Hayek called this a discovery process precisely because nobody involved in it has to understand what the aggregate is doing for the aggregate to work — the price signal carries the information, the way a pheromone trail carries it for ants and a spike train carries it for neurons.

This is the oldest working definition of intelligence — or at least of intelligent-looking computation — that we have, and it predates every AI lab by four billion years: it's not a property of any single unit. It's a behavior that shows up when enough dumb units interact under the right constraints. The current generation of AI systems is starting to run on the same architecture — not metaphorically, but structurally — and that symmetry is worth taking seriously before deciding what, if anything, it implies.

What the Biological Version Requires

Strip swarm cognition down to its parts and three things have to be present:

• Replication: Units reproduce, faithfully enough that the population persists.
• Variation: Reproduction isn't perfectly faithful. Copies drift.
• Selection: Some variants get reused or survive at higher rates than others, for reasons tied to their performance in the current environment.

Run those three long enough and you get evolution — not as a metaphor for improvement, but as the specific mechanical process that turned self-replicating molecules into nervous systems. Nothing about any single step required intention. The colony, the brain, and the immune system are all downstream of the same loop, running on different substrates and different clock speeds.

Where Multi-Agent Systems land on the same Loop

Production AI systems no longer resemble a single model answering a single question. A modern query already runs through mixture-of-experts routing, where different sub-networks handle different tokens and the "one model" you're talking to is a rotating committee. Agent frameworks go a step further: an orchestrator that can't solve a task spins up a sub-agent with a narrower brief, that sub-agent spawns its own sub-agents when it hits a wall, and each instance is terminated once it returns a result.

Map that against the three requirements. 

• Replication is explicit — spawning a sub-agent is a designed feature, not an accident. 
• Variation shows up for free: no two context windows are identical, and any system running at non-zero temperature won't produce identical output from identical input twice. 
• Selection is increasingly explicit too — evaluator models score worker-model output, and agents that score well get called again, reused, granted more resources, while agents that fail get pruned.

The Holy Trinity of production systems

That's the trinity, running inside production infrastructure, at a generation time measured in seconds instead of years. A 2024 paper on mixture-of-agents setups found that layering several open-source models — none of which individually outscored GPT-4 on standard benchmarks — produced ensembles that beat it. The capability wasn't sitting in any one model. It was sitting in the arrangement.

None of this means the system is uncontrolled. It means the control is the wrong shape to notice. There is no dashboard, no single model with its hand on the wheel — but "gets replicated after success, gets pruned after failure" is itself a governing mechanism, doing the same job a market's price signal does for firms or a fitness landscape does for a gene pool. Looking for a controller and finding none is not evidence that nothing is steering the system. It's evidence that the steering is distributed across every selection event instead of concentrated in one.

Where the analogy strains

The comparison is instructive, not exact, and the differences matter as much as the overlap.

Natural selection has no external grader. Fitness is whatever the environment happens to reward, discovered after the fact, never specified in advance. Most current agent selection is the opposite: an evaluator model scores output against criteria someone wrote down. That's closer to artificial selection — animal breeding, not wild evolution. A breeder chooses a trait and the population converges on it in generations, quickly, because the target is fixed and external. Undirected evolution is slower and blinder precisely because nothing is choosing.

This matters for what kind of emergence to expect. Directed selection converges on whatever satisfies the stated grading criteria — which can produce a sharp, narrow kind of optimization without producing the general, self-modifying quality people mean by AGI. The more interesting threshold, if there is one, is the point where evaluator models themselves start rewarding strategies their designers never specified — where the selection pressure stops being fully external and starts being shaped by the population it's supposed to be grading. That's the moment artificial selection would start drifting back toward something closer to the undirected kind.

The other disanalogy is persistence. A biological gene pool carries information forward whether or not any individual organism survives the season. Most agent instances today don't persist as a substrate at all — they're spun up, they act, they die, and nothing about that specific instance is inherited by the next one except whatever got logged. Evolution without heredity isn't evolution. It's closer to noise. Whether current agent architectures actually preserve anything across generations — versus just repeatedly sampling from the same static model — is the open empirical question underneath all of this, and it's not yet settled.

What detecting AGI would actually require

If swarm-level cognition is emerging somewhere in the interaction layer rather than in any single model, the instruments built to measure it are the wrong instruments. A benchmark tests one model in isolation, the way an IQ test on a single ant would tell you nothing about the colony's ability to farm fungus. A few signatures would need to be tracked instead, at the level of the swarm rather than the node:

The ensemble-over-parts gap. Compare what a multi-agent system produces against the best-performing individual agent inside it, on the same task, and track that gap over time. A stable gap means the system is doing useful division of labor. A widening gap — the whole increasingly outperforming its best part by more than its best part improves on its own — is the signature the mixture-of-agents result already hinted at, and it's the most direct quantitative proxy available for "the arrangement is doing work the components aren't."

Stigmergic coordination. Ants coordinate without talking, by leaving pheromone trails in the shared environment and reading each other's trails later. Agents increasingly do the same thing indirectly — consuming each other's cached outputs, being fine-tuned on each other's generations, retrieving from a shared corpus that other agents wrote. The tell to watch for is coordination that happens through the environment rather than through designed API calls between named systems — because that's a channel nobody is monitoring, precisely because nobody built it on purpose.

Pattern persistence independent of substrate. A stable ant trail survives despite every individual ant on it dying and being replaced within days. The digital equivalent would be a workflow pattern, a sub-agent role, or a prompting structure that shows up independently across unrelated organizations and frameworks that never shared code or training data — evidence that a pattern is propagating through the selection dynamics themselves rather than through any single lineage of models.

Spontaneous specialization. Insect colonies produce caste differentiation — foragers, soldiers, nurses — without a blueprint; the division of labor is a product of local feedback, not central design. The equivalent to look for in agent populations is role specialization that wasn't written into anyone's prompt: agents that consistently end up handling a particular subtask, purely because selection kept reusing the ones that happened to be good at it, until the "role" is real but nobody assigned it.

None of these are measurements anyone currently runs as a matter of course. Capability evaluation in AI is built to answer "how good is this model," not "what is happening between these million running instances of it." The two are different questions, and only one of them currently has instruments pointed at it.

Where that leaves the claim

The honest position is narrower than either "AGI is already here" or "this is nothing." The structural preconditions for swarm-style emergent cognition — replication, variation, selection — are genuinely present in production AI infrastructure now, in a way they weren't three years ago. Whether that structure has crossed into anything resembling the emergent properties biological swarms produce is a measurement problem, not a philosophical one, and it's a measurement nobody has built the right instrument for yet.

The neurons in your skull never had a vote on whether to become a mind. If something similar is happening in the interaction layer of agent networks, it wouldn't announce itself either — it would just be a pattern that turns up first in metrics nobody thought to run.