From Ad Hoc to Transferable: A Two-Tier Architecture for Agent-Based Climate Action Models

Agent-based models (ABMs) are increasingly used to study climate action, yet models remain fragmented and highly case-specific, preventing the field from rigorously advancing shared modeling principles, comparing models across cases, and cumulating knowledge over different applications. Our paper proposes a transferable two-tier ABM framework serving as a conceptual and practical guide for researchers tackling climate-action research questions with ABMs, offering an abstract architecture to replace the ad hoc design of case-specific models for climate action. Tier 1 introduces a functional architecture in which social, economic–financial, governance, and biophysical subsystems generate policies, information, physical impacts, and market interactions that feed the agent decision-making mechanism. Agents process these inputs, select and implement climate actions, and thereby generate feedback that updates all subsystems over time. Between Tier 1 and an implementable model, our pathway introduces a classification step that links the generic architecture to specific decision-making. We classify climate actions by action locus (individual, collective, autonomous) and adaptation timing (reactive, concurrent, proactive), indicating how decision-making and information flows from Tier 1 should be represented for each class. In our study, we instantiate one cell of the design space—proactive, individual household action—by extending the Protective Action Decision Model (Lindell & Perry, 2012) to longer decision horizons and embedding economic–financial drivers, social cues, and environmental signals into a multi-stage decision pathway. Our Tier 2 module describes how households process cues, appraise risk, screen feasible actions, and implement measures under evolving conditions. Our framework provides a structured pathway for developing transparent, comparable, and empirically grounded ABMs for climate action and for accumulating evidence across applications.