Separate Windows: Ongoing Independent Grounding as a Condition for Open-Ended AI Collectives

Collectives of AI agents are increasingly proposed as a route to open-ended learning and as a check on any single model's errors. The current work argues that such a collective stays stable and open-ended over time only as far as each member keeps its own, ongoing contact with a changing world. The overlap among agents' channels to the world is called grounding correlation herein. As it rises, false consensus rises and diversity of belief falls. Differences in model, prompt or process also decorrelate errors, but they cannot tell a collective about changes in the world that none of its members have observed. The underlying idea is established in collective epistemics and cultural evolution, from Rogers' paradox to wisdom-of-crowds and network-epistemology results; the contribution is carrying it into AI collectives as an architectural requirement. Preliminary agent-based simulations, deliberately minimal, support the claim. When agents stopped observing the world, the collective's accuracy fell to chance within one or two world changes, and independent windows by themselves gave no advantage without ongoing observation. With ongoing observation, independence kept agreement trustworthy: when 90% of agents agreed, they were wrong 3.5% of the time with independent windows but 28.5% once windows overlapped by half, because learning from one another can turn partly shared blind spots into confident group error. Across the simulations, four primary variables drove the outcomes: ongoing access to the world, independent access, model diversity and the grounding rate (how often agents observe the world rather than copy one another); the first three together kept a full collective stable over long runs. Agent dispositions, such as how agents copy and how strongly they defer to the majority, shifted the points at which effects appeared, but no disposition made up for losing any of the first three. The grounding rate also supplies a boundary condition: the effect holds where agents observe often enough to overturn an outdated consensus. The results also carry a safety implication: agreement among agents that share a window is weak evidence that what they agree on is correct.