Self-Forming Agents: Closed-Loop Individuation from Identical Copies

Artificial agents are usually treated as interchangeable instances of a trained design. This paper asks whether identical agents can become non-interchangeable through closed-loop histories, without external reward or retraining a shared base model. We introduce a minimal Self-Forming Agent (SFA) with three online fields: recent raw history (R), path-stabilized history (C), and cue–action relations (W). W predicts an intrinsic consequence defined as relational change multiplied by historical appearance in C. Three synthetic experiments separate mechanism from task competence. Across 12 paired seeds, product-based W selected product-optimal actions more often than no-W, MA-only, additive, frozen, and temporally shuffled controls; its contribution beyond change alone appeared when the factors favored different actions. Next, 32 exact twin pairs were forked into different cue–action histories. On identical reunion probes, twins assigned 0.984 probability to their own mappings and differed by 0.982 total variation. Revaluation changed current choice while preserving consolidated lineage, and W transplantation switched every divergent cue choice to the donor mapping. Finally, ten clones were followed in each of 12 populations. Independent sampling produced no persistent difference when agent state was frozen, but yielded mean pairwise total variation of 0.240 when actions changed the world and consequences were written back. After 1,600 identical exposures, 98.8% of behavioral difference remained; W transplantation transferred 84.9% of donor proximity. These results establish a proof of mechanism for history-dependent individuation, not general intelligence or consciousness. SFA is a compact adaptive layer whose present dispositions are continually formed by enacted history.