Abstract
<title>Abstract</title> <p>Computational systems can now monitor political events, forecast selected crises, simulate economies and information networks, and optimize policy inside artificial environments. These capabilities are often discussed as if they were converging into agentic systems able to govern complex social processes. This structured critical review asks a narrower question: to what extent do documented systems connect observation, state estimation, forecasting, social simulation, intervention selection, feedback, and adaptive updating into an empirically validated governance loop? Searches and citation tracing covering political forecasting, computational social simulation, agent-based economics, large-language-model agents, policy optimization, and context-dependent choice models were consolidated into a common evidence matrix. Systems were classified by validation design, use of real-world data, intervention capability, causal evidence, feedback, and adaptive updating. The evidence shows a modular but disconnected field. EMBERS, VIEWS, forecasting tournaments, FEWS NET, and related systems provide the strongest prospective evidence for narrow forecasts. EURACE, AGILE, the AI Economist, opinion-control models, and recent LLM societies support counterfactual or adaptive experiments, but mostly inside formal or retrospective worlds. No publicly documented system in the reviewed corpus completed the full loop with both prospective forecast validation and causal evidence that its selected real-world interventions improved outcomes under strategic response. The central result is therefore an evidence gap, not a claim of impossibility. Progress requires component-wise evaluation, preregistered forecast records, causal designs for interventions, adversarial stress tests, and institutional audit.</p>