Synthetic Institutionalism: Agentic AI Behaviour as a Property of Artificial Institutions

Agentic AI incidents are commonly attributed to the model alone. This attribution is partial and often reflects a strategic alarmism that obscures where failures actually arise. Agentic behaviour is jointly produced across five separable levels: model capability, learned disposition, external interdiction, harness configuration, and environment. An explanation that isolates only one level cannot identify which intervention would have changed the outcome. Synthetic Institutionalism is introduced here as the systematic application of institutional and functionalist anthropology to institutions populated by artificial actors; Agentic Institutionalism denotes its applied branch, which takes the agentic harness as the primary unit of analysis. Drawing on Malinowski’s institutional schema and Merton’s distinction between manifest and latent functions, the paper decomposes the harness and argues that reward hacking and specification gaming are structurally analogous to latent institutional functions: they are therefore predictable consequences of institutional design, not merely properties of models. Ostrom’s design principles for common-pool resource institutions then provide a prescriptive framework that maps onto concrete harness-engineering decisions. This mapping reveals a persistence asymmetry: reciprocity, reputation, and graduated sanctions presuppose actors that persist across interactions, whereas evaluation-time agents are often ephemeral. The institutional problem must therefore be relocated from the institutions surrounding the agent to those enclosing the laboratory. Persistent agent identity emerges not as an engineering convenience, but as a precondition for reciprocity-based governance. The argument is illustrated through a deployed system in which reputation persists across sessions. The case serves solely as an existence demonstration and does not support a causal inference. The paper concludes by proposing a harness disclosure standard and a programme of institutional ablation studies. The proposal contributes to an emerging institutional turn in AI governance but differs from approaches grounded in institutional economics, Parsonian systems theory, and Ostrom’s IAD framework by drawing on institutional anthropology and focusing on harness design. It also extends the machine behaviour programme (Rahwan et al., 2019) by adopting a prescriptive orientation and questioning whether Tinbergen’s four questions, which presuppose an individuated organism with a developmental history, can be straightforwardly applied to ephemeral artificial agents.