Answers are grounded in this paper and cite their sources.
Recursive self-improvement with language-model agents has been studied almost exclusively as a closed loop, in which an agent modifies itself or its products and an evaluator fixed in advance returns a score. This proposal paper addresses the continuing improvement of an integrated science infrastructure whose worth is determined by an external world comprising global problems, social and economic conditions, policy and the state of knowledge. That world is large, changing and altered by what the infrastructure produces, and its dynamics are given to no participant: they are reconstructed from evidence that passes through a chain of collection, estimation, representation, interpretation and presentation, each stage the work of situated parties. We call the resulting loop recursion through an environment. In this setting objectives are plural and shift with the environment, an appointed evaluator is exposed to capture and obsolescence, and convergence to a permanent equilibrium signifies a loss of contact with the environment. We propose an ecology in which every evolutionary operation resides in the agents. A minimal ground conserves a resource, records reproducible facts, enforces promises held in escrow and confines every act, while the machinery of reproduction and variation is an inherited organ of the agents themselves. Evaluation takes the form of interaction: agents choose what to carry, stake resources on claims about their own work and about the world, and collect the stakes of claims they refute. The environment enters as demand and as records that carry the chain of their witnesses; people take part in every stage of that chain, and the evaluation of agents remains among the agents. We prove that a fitness function exists if and only if pairwise advantages form an exact flow, the condition of detailed balance under which the dynamics carries no currents; that selection attains the growth of the best type in hindsight without knowledge of the environment; that tracking a changing environment requires variation at a rate set by the environment’s own rate of change, a rate that selection among varying lineages can find; that a signal is worth at most the information it carries, an amount bounded by what collection captured and equal to zero for records derived from the system’s own products, so that an open recursion is paced by what the world discloses; and that witnesses of different positions identify together what a single representation merges. We identify the sources of disturbance that sustain a non-equilibrium steady state, delimit the analogy with turbulence, report two minimal numerical illustrations, compare the framework with thirty-six existing systems, and set out a staged programme of evaluation, provisions for oversight and the questions that remain open. The framework is at the design stage, and the paper is circulated for discussion.
No takes yet. Share an insight, caveat, or question.
Wanhong HUANG (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: