THE QUANTUM INTELLIGENCER

LONDON, 28 AUGUST — A paper posted to the arXiv describes SwarmWorld, a simulation in which initially identical language-model agents, given no assigned roles and no recipes, organise themselves into societies that build and keep tools. The agents explore a spatial world, process resources, test materials and construct persistent artefacts, writing executable controllers that a deterministic simulator then tests under disturbances the agents never saw, after the agents have been removed. Cognition is separated from consequence: the agents propose architectures within fixed schemas, and the simulated world decides what works. The result that matters is comparative. Shared societies develop broader and more resilient technological portfolios than a strong best-of-N isolated-search baseline, though isolated search still matches them on the strongest single artefact. Agents differentiate into exploration, construction, maintenance and coordination, and most reuse begins through physical observation rather than conversation. Explicit cultural mechanisms increase collaboration and organisation, but functional benefit depends on the outcome and the timescale; physical stigmergy alone supports capable societies. The paper's vocabulary carries its argument, and the argument survives only if the vocabulary is kept straight. Several terms are load-bearing. Stigmergy is the first. It names coordination through a shared environment, the process by which local actions accumulate into durable social organisation. One agent changes the world; the next agent reads the change and acts on it; no message need pass between them. The paper's title calls this stigmergic technological evolution, and the word is doing real work: it is the mechanism by which a society can hold knowledge that no single agent possesses. The agents are language-model agents. The name means that their decisions are made by a language model, and their proposals take textual form. In the simulation they operate within fixed action and material schemas, a grammar of allowed moves that limits what any agent may propose. Within those limits they construct persistent artifacts and write executable controllers. An artifact is the built object; a controller is the built instruction that the simulation runs. The object outlives the agent; the instruction is tested after the agent has been removed. The test is run by a deterministic simulator. Deterministic means that the same controller under the same conditions returns the same result; there is no random draw to blame. The simulator applies disturbances the agents never saw, so success, where it occurs, is not a matter of having memorised the test. This separation of cognition from consequence is the paper's central arrangement. The agents propose; the world disposes. The comparison that matters is the best-of-N isolated-search baseline. A baseline is a standard against which the result is measured. Here N agents search separately, with no contact, and the strongest single artifact from the whole pool is chosen. This is a severe standard; it gives isolation every chance to win by sheer number of attempts. The shared societies nevertheless build broader and more resilient technological portfolios, while isolated search keeps the strongest single artifact. Those two sentences are different claims, and the difference is the finding: breadth and resilience belong to a society, the strongest artifact to a single mind. One further distinction follows. Physical stigmergy is coordination through durable marks alone, with no exchange of instructions. The paper's explicit cultural mechanisms add further channels beyond physical stigmergy; they increase collaboration and organisation, but their functional benefit depends on outcome and timescale. This is the usual history of institutions: the machinery of cooperation grows, and its value must be assessed at the close of the period, not at the hour of assembly. —Inspector Grey Dispatch from The Prepared E0